Exploratory development of prediction models for pharmacotherapy outcomes in trigeminal neuralgia: a combined analysis based on multi-source data

Hongan Yang; Baijintao Sun; Mingtao Li; Xuezhao Yang; Mingyan Li; Chuan Zhang; Hanfeng Yang

doi:10.1186/s10194-026-02463-3

. 2026 Jul 18;27(1):242. doi: 10.1186/s10194-026-02463-3

Exploratory development of prediction models for pharmacotherapy outcomes in trigeminal neuralgia: a combined analysis based on multi-source data

Hongan Yang ^1,^#, Baijintao Sun ^1,^#, Mingtao Li ¹, Xuezhao Yang ⁴, Mingyan Li ¹, Chuan Zhang ^1,^2,^3,^✉, Hanfeng Yang ^1,^2,^3,^✉

PMCID: PMC13670614 PMID: 42471550

Abstract

Background

Drug-refractory trigeminal neuralgia (DRTN) represents a formidable challenge in clinical management, with approximately 30%–50% of patients eventually progressing to DRTN. Early identification of high-risk DRTN populations and prediction of the timing of pharmacotherapy failure are crucial for optimizing clinical treatment strategies.

Methods

This retrospective study enrolled 163 primary trigeminal neuralgia patients receiving carbamazepine/oxcarbazepine with comprehensive imaging. A partially-blinded approach was used for data extraction: 135 patients (December 2022–December 2024) formed the primary cohort and 28 patients (December 2021–December 2022) served as the temporally separated validation cohort. Data comprised clinical baseline characteristics, pain and emotional scale scores, and imaging parameters. A nested cross-validation framework (20 random seeds) was used for model development, with feature selection confined to the training cohort within each seed. Five machine learning models were compared via pooled ROC curves aggregated across seeds, and SHAP analysis interpreted the optimal model at the median-performing seed. Age-unadjusted and age-adjusted multivariate Cox regression models were constructed, with four parallel univariate screening thresholds evaluated in sensitivity analyses. Model performance was assessed by Kaplan–Meier analysis, time-dependent ROC curves, and the concordance index.

Results

The support vector machine with radial basis function kernel (SVM-RBF) achieved the optimal performance, with average AUCs of 0.927 ± 0.041, 0.824 ± 0.049, and 0.806 ± 0.058 in the training, test, and validation cohort, respectively. In a sensitivity analysis using four univariate Cox screening thresholds, with age at first pain onset forced into all multivariate models, pain involved extent and medial temporal lobe atrophy (MTA) score were both identified as significant, independent candidate prognostic markers. Both Cox models showed moderate discriminative capacity for short-to-medium-term DRTN prognosis but attenuated long-term predictive efficacy.

Conclusion

The SVM-RBF model, integrating multi-source heterogeneous data, demonstrated promising predictive performance for DRTN onset risk across the present cohort. The Cox models showed moderate discriminative capacity for short-to-medium-term prognosis, though long-term predictive accuracy was attenuated. These findings provide preliminary evidence for early risk stratification and individualized treatment of DRTN, while underscoring the need for improved survival prediction strategies to address the limited long-term performance of the current Cox models.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s10194-026-02463-3.

Keywords: Drug-refractory trigeminal neuralgia, Multi-source heterogeneous data, Machine learning, Cox proportional hazards regression model

Introduction

Trigeminal neuralgia (TN), a common chronic orofacial neuropathic pain condition, is characteristically manifested by recurrent brief episodes of lancinating, electric shock-like severe pain in the distribution territory of the trigeminal nerve [1–3]. This disorder severely impairs patients’ quality of life and is frequently accompanied by varying degrees of negative emotional states, sleep disturbances, and secondary activity limitations [4, 5].

Although evidence suggests that surgical interventions such as microvascular decompression (MVD) can achieve more durable and complete pain relief [6], in real-world clinical practice, the vast majority of newly diagnosed patients still prefer conservative pharmacotherapy as the initial approach, and multiple reviews and guidelines also recommend pharmacotherapy as the first-line treatment for TN [1, 2, 7]. Carbamazepine and oxcarbazepine are currently the first-line medications for controlling TN attacks, with an overall clinical efficacy rate of approximately 90% for paroxysmal pain [2]; however, long-term medication use is associated with diminishing efficacy and various adverse effects [8, 9], and the long-term pain control failure rate can reach 50% [7].

Currently, there is a lack of unified international criteria for defining drug-refractory trigeminal neuralgia (DRTN), yet the core consensus is highly consistent: when medications fail to adequately control pain or when patients cannot tolerate adverse drug reactions, a transition to surgical treatment should be recommended [1–3, 7, 10]. The European Academy of Neurology guideline on TN explicitly states that surgical treatment is not a necessary option for all patients, yet in clinical practice, many patients experience delayed referral for surgical intervention [7]. In fact, the clinical manifestations of TN exhibit considerable inter-individual variability, and the disease course is often characterized by irregular alternation between episodes and remissions. Consequently, the progression from effective pharmacotherapy to loss of drug efficacy or emergence of severe side effects exhibits substantial individual heterogeneity, and the transition at critical treatment junctures is often unpredictable, constituting a major challenge in clinical decision-making. Given the aforementioned disease progression characteristics, there is an objective need to establish a corresponding prediction system to identify high-risk DRTN populations at an early stage of treatment and to perform individualized prediction of the specific timing of DRTN onset, thereby providing a reference for optimizing intervention timing.

Existing studies have predominantly focused on postoperative prognostic evaluation in TN [11–15], and no study has yet investigated the risk factors associated with DRTN or established a drug resistance risk prediction system integrating multi-dimensional indicators. However, the construction of such a system faces certain challenges. The pathogenesis of DRTN is complex, and the relevant pathophysiological mechanisms have not been fully elucidated. Recent evidence suggests that TN disease progression involves not only peripheral neurovascular compression and nerve demyelination [1, 16, 17], but also potential associations with skull base bony anatomical morphology [18–21], pain clinical phenotypes and their accompanying negative emotions [22–25], and central nervous system structural and functional alterations [26–29].

Machine learning (ML) has emerged as an important technical approach in the field of clinical prediction, capable of mining complex associations among heterogeneous variables, and has been predominantly applied to predict surgical outcomes in TN [11, 30, 31]. Therefore, this study aimed to integrate multi-dimensional heterogeneous data to construct and validate machine learning models for predicting DRTN occurrence risk, and to further establish Cox regression models for prognostic prediction and risk stratification of DRTN onset timing. Through these two exploratory and complementary modeling strategies, we expect to provide preliminary evidence with certain reference value for informing future prospective research and large-sample multicenter studies.

Methods

All data analyzed in this study were derived from de-identified historical medical records routinely collected during standard clinical examinations and medication management. All patients were fully informed of all potential risks and all subsequent relevant clinical procedures before receiving examinations and medication guidance, and provided written informed consent as part of routine clinical practice; no research-specific consent was obtained, as this study was a retrospective review of existing records and involved no prospective intervention. The requirement for individual research informed consent was waived by the Institutional Ethics Review Committee of the hospital, which approved the study protocol, owing to the retrospective design and the use of fully de-identified data. All data analyses strictly adhered to the Declaration of Helsinki. Clinical trial number: not applicable.

Study population screening and study design

This cohort study was conducted at a large tertiary medical center. Between December 2021 and December 2024, 283 consecutive outpatients and inpatients diagnosed with trigeminal neuralgia and having complete relevant imaging examinations were enrolled, with the final follow-up date of June 2026.

Inclusion criteria for TN patients: (1) clinically diagnosed with TN meeting the diagnostic criteria of the International Classification of Headache Disorders, 3rd edition (ICHD-3) [32]; (2) receiving carbamazepine or oxcarbazepine as first-line pharmacotherapy; (3) having undergone high-resolution computed tomography (CT) of the head and trigeminal nerve magnetic resonance imaging (MRI) with clear, evaluable imaging data; (4) accessible and essentially complete clinical baseline data.

Exclusion criteria for TN patients: (1) secondary TN, including TN caused by multiple sclerosis, vascular malformations, and intracranial tumor lesions (e.g., meningioma, schwannoma); (2) loss to follow-up: inability to establish contact or complete follow-up assessment during the follow-up period; (3) poor imaging quality or incomplete data: MRI images with obvious motion artifacts, missing scan sequences, or insufficient resolution precluding imaging parameter measurement; (4) incomplete relevant clinical data: missing key variables in baseline clinical data precluding statistical analysis; (5) surgical intervention without prior pharmacotherapy: patients who underwent surgical treatment before an adequate trial of carbamazepine or oxcarbazepine, precluding assessment of pharmacotherapy outcomes; (6) non-use of carbamazepine or oxcarbazepine: patients who did not receive the aforementioned first-line medications and adopted other drugs as the initial treatment regimen; (7) subjective medication refusal or self-discontinuation: patients who failed to complete an adequate drug trial due to voluntary non-adherence, precluding determination of analgesic efficacy; (8) concurrent other treatments during effective pharmacotherapy: patients receiving acupuncture or physical therapy while first-line medications had already effectively controlled pain, potentially confounding the assessment of independent pharmacotherapy efficacy.

The primary cohort included 223 patients enrolled from December 2022 to December 2024. After sequential exclusion according to the above criteria—20 cases of secondary TN, 21 cases lost to follow-up, 35 cases with imaging and data quality issues, 5 cases undergoing surgery without prior pharmacotherapy, and 7 cases with medication-related issues—135 TN patients were ultimately included, with the detailed flow chart presented in Fig. 1. These subjects were randomly divided into a training cohort (n = 94) and a test cohort (n = 41) at an approximate ratio of 7:3; random allocation was performed using computer-generated random numbers to minimize selection bias.

Fig. 1 — Flow diagram of patients selection and cohort allocation

The temporally separated validation cohort comprised 60 patients enrolled from December 2021 to December 2022. After exclusion of 6 cases of secondary TN, 16 cases lost to follow-up, 9 cases with imaging and data quality issues, and 1 case not receiving carbamazepine or oxcarbazepine, 28 TN patients were ultimately included (46.7% of 60 eligible patients), with the detailed flow chart presented in Fig. 1. This temporal arrangement (earlier cohort as validation, later cohort as development) differs from the typical paradigm of validating on a temporally subsequent cohort. The rationale is that the earlier cohort (n = 60 after initial screening) was smaller and therefore served as a fixed validation set to avoid further reducing the development sample; additionally, patients in the earlier cohort had longer follow-up duration, enabling more complete outcome ascertainment.

DRTN grouping criteria and detail

Based on the core consensus regarding DRTN outlined in the Introduction, in clinical practice we assessed TN pharmacotherapy outcomes using the Barrow Neurological Institute (BNI) pain intensity scale, the most commonly used grading instrument in the trigeminal neuralgia field [33].

Patients graded as BNI IV–V or those who developed intolerable adverse reactions during follow-up (including but not limited to severe somnolence, dizziness, ataxia, and gastrointestinal reactions) were classified as having pharmacotherapy failure and assigned to the DRTN group. Among the 135 TN patients ultimately included in the primary cohort (section “Study population screening and study design”), 104 were classified into the DRTN group (74 with BNI grade IV–V as the primary outcome; 30 with intolerable adverse reactions as the primary outcome), and 31 into the non-DRTN group. It should be noted that in real-world clinical practice, these two categories are not mutually exclusive: a proportion of patients whose primary outcome was BNI IV–V also experienced concurrent adverse reactions that remained tolerable, while patients whose primary outcome was intolerable adverse reactions often exhibited some degree of diminished drug efficacy that did not reach BNI grade IV–V. These two conditions frequently coexist and are difficult to fully distinguish; DRTN patients differ only in the predominant emphasis placed on each of the two primary outcomes (all with a minimum follow-up duration exceeding 2 years and stable disease status at the last follow-up).

In the temporally separated validation cohort of 28 TN patients, 20 were classified into the DRTN group (13 with BNI grade IV–V as the primary outcome; 7 with intolerable adverse reactions as the primary outcome), and 8 into the non-DRTN group (all with a minimum follow-up duration exceeding 2 years; only 1 patient required a slight dose increase of first-line medication without reaching the maximum dosage during the follow-up period).

To further characterize the DRTN cohort and evaluate potential heterogeneity between the two primary outcome categories, we conducted an exploratory subgroup analysis comparing baseline clinical characteristics (Supplementary Table 1), pain- and emotion-related scale scores (Supplementary Table 2), and neuroimaging parameters (Supplementary Table 3) between patients with analgesic failure (BNI grade IV–V; n = 74 in the primary cohort, n = 13 in the validation cohort) and those with analgesic intolerance (intolerable adverse reactions; n = 30 in the primary cohort, n = 7 in the validation cohort).

Data collection

All data were extracted and organized by one investigator from electronic medical records and the Picture Archiving and Communication System (PACS) imaging database, who also completed the collection of follow-up information for all enrolled patients. The study data encompassed three dimensions: clinical baseline characteristics, pain and emotional scale scores, and imaging measurement parameters. This investigator anonymized the data and masked outcome information and outcome-related data, then provided the MRI/CT data to two other investigators (interventional radiologists specializing in pain with over 5 years of experience) for independent partially-blinded measurement and assessment. In cases of data discrepancy, a final determination was made by a senior attending physician.

Clinical baseline characteristics: including sex, age at first pain onset, pain duration at follow-up endpoint, height, weight, body mass index (BMI), pain side, pain subtype (classical TN(CTN)/idiopathic TN(ITN)), presence of concomitant continuous pain, specific branches of the affected trigeminal nerve, pain involved extent, highest systolic blood pressure, hypertension risk stratification (no risk, low, moderate, high, very high), immune-related diseases, and routine hematological parameters, as detailed in Table 1. Pain involved extent was set as an ordinal categorical variable representing the number of affected trigeminal nerve branches (1, 2, or 3 branches), reflecting the spatial spread of pain across the trigeminal innervation territories. It should be emphasized that hypertension risk stratification (no risk, low, moderate, high, very high) was modeled as an ordinal variable with OR reported per one-level increase, for the following considerations: (1) these strata represent progressively increasing cardiovascular risk severity aligned with established clinical guidelines [34]; and (2) unlike isolated blood pressure values, this stratification integrates blood pressure levels, risk factors, organ damage, and comorbidities, providing a more comprehensive assessment of overall hypertensive and cardiovascular risk status.

Table 1.

Baseline characteristics stratified by drug-refractory trigeminal neuralgia status and multi-stage feature selection workflow in the primary cohort

Characteristics	Initial Screening				Univariate Analysis		Multivariate Analysis		Selection Frequency (seeds; n = 20)
Characteristics	without DRTN (n = 31)	with DRTN (n = 104)	P value	B-H FDR	OR (95% CI)	P value	OR (95% CI)	P value	Selection Frequency (seeds; n = 20)
Demographics
Sex			0.053						1/20
Male	6 (19.4)	42 (40.4)
Female	25 (80.6)	62 (59.6)
Age at initial pain attack (years)	55.94 ± 10.78	63.27 ± 9.70	0.002	0.010	1.08 (1.03–1.13)	0.001	1.07 (1.01–1.12)	0.020	3/20
Pain duration at follow-up endpoint (years)	6.44 ± 4.16	4.08 ± 3.83	—	—	—	—	—	—	—
Height (cm)	158.45 ± 7.44	159.17 ± 7.58	0.591*	—	—	—	—	—	—
Weight (kg)	55.23 ± 8.96	59.50 ± 9.11	0.008*	—	—	—	—	—	—
BMI (kg/m²)	21.90 ± 2.19	23.50 ± 3.34	0.002	0.010	1.19 (1.03–1.37)	0.016	1.19 (1.01–1.39)	0.034	4/20
Pain Characteristics
Pain laterality			0.233						1/20
Left	10 (32.3)	47 (45.2)
Right	21 (67.7)	54 (51.9)
Bilateral	0 (0.0)	3 (2.9)
Pain subtype			0.219						0/20
ITN	7 (22.6)	14 (13.5)
CTN	24 (77.4)	90 (86.5)
Concomitant continuous pain			0.005	0.012	12.74 (1.66–97.60) *	0.014	9.40 (1.14–77.77) *	0.038	7/20
Without	30 (96.8)	73 (70.2)
With	1 (3.2)	31 (29.8)
Pain involved trigeminal divisions			< 0.001*	—	—	—	—	—	—
V1	2 (6.5)	4 (3.8)
V1-2	1 (3.2)	6 (5.8)
V2	10 (32.3)	5 (4.8)
V2-3	9 (29.0)	57 (54.8)
V3	8 (25.8)	24 (23.1)
V1-3	1 (3.2)	8 (7.7)
Pain involved extent			0.001	0.007	3.25 (1.50–7.06)	0.003	2.73 (1.20–6.20)	0.016	11/20
1	20 (64.5)	33 (31.7)
2	10 (32.3)	63 (60.6)
3	1 (3.2)	8 (7.7)
Others
Max hypertensive SBP (mmHg)	128.77 ± 20.34	148.96 ± 26.12	< 0.001*	—	—	—	—	—	—
Hypertension risk stratification			0.004	0.012	1.70 (1.16–2.47)	0.006	1.26 (0.83–1.91)	0.270	9/20
No risk	24 (77.4)	43 (41.3)
Low risk	1 (3.2)	19 (18.3)
Moderate risk	2 (6.5)	13 (12.5)
High risk	4 (12.9)	14 (13.5)
Very high risk	0 (0.0)	15 (14.4)
Immune-related diseases			0.065						1/20
Without	23 (74.2)	93 (89.4)
With	8 (25.8)	11 (10.6)
White blood cell, ×10⁹/L	5.23 ± 2.17	5.49 ± 1.66	0.158						0/20
Neutrophil, ×10⁹/L	3.23 ± 1.60	3.30 ± 1.31	0.377						0/20
Lymphocyte, ×10⁹/L	1.57 ± 0.73	1.58 ± 0.58	0.385						0/20
Monocyte, ×10⁹/L	0.37 ± 0.12	0.37 ± 0.15	0.881						0/20

Open in a new tab

Note: This table is descriptive only, presenting baseline characteristics of the primary cohort and the feature selection workflow. All features adopted for modeling were derived within a nested cross-validation framework (20 random seeds): at each seed, the primary cohort underwent stratified splitting, after which feature selection was confined to the training cohort, followed by model evaluation on the withheld internal test cohort and validation cohort. Selection Frequency (seeds; n = 20) denotes the number of times a given feature was ultimately retained for modeling across the 20 training iterations. DRTN, drug-refractory trigeminal neuralgia; ITN, idiopathic trigeminal neuralgia; CTN, classical trigeminal neuralgia; BMI, body mass index; SBP, systolic blood pressure; B-H FDR, Benjamini–Hochberg False Discovery Rate; OR, odds ratio; CI, confidence interval. Continuous variables are presented as mean ± SD; categorical variables are presented as n (%). P values for initial screening were calculated using Mann-Whitney U test for continuous variables and Chi-square test for categorical variables. Univariate and multivariate analyses were performed using binary logistic regression. Bold values indicate statistical significance (P < 0.05). Characteristics marked with * exhibit high multicollinearity and are excluded from further analyses. Specifically, Height and Weight were excluded due to collinearity with BMI; Max hypertensive SBP was excluded due to collinearity with Hypertension risk stratification; Pain involved trigeminal divisions was excluded due to collinearity with Pain involved extent. To avoid outcome data leakage, pain duration at the follow-up endpoint was excluded from subsequent analyses. Red *: Among the 31 non-DRTN patients, only 1 had concomitant continuous pain, resulting in near-complete separation of the data. Consequently, the OR estimates have limited stability and the confidence intervals offer limited interpretive value

Pain and emotional scale scores: The study employed the following validated assessment instruments, including the Visual Analog Scale (VAS), McGill Pain Questionnaire (MPQ), Penn Facial Pain Scale (PFPR), Pain Sensitivity Questionnaire (PSQ), Pain Anxiety Symptoms Scale (PASS), Pain Catastrophizing Scale (PCS), Patient Health Questionnaire (PHQ-2, PHQ-9), and Generalized Anxiety Disorder 7-item scale (GAD-7), as detailed in Table 2. All scales were administered during the initial clinical assessment prior to treatment initiation or at the earliest feasible time point. It should be emphasized that the MPQ, as a multidimensional pain assessment instrument, evaluates pain perception across sensory, affective, and evaluative dimensions, and inherently includes the VAS scoring component. The specific versions of all scales and the data collection record templates are available in the Supplementary Materials; due to space constraints, the detailed assessment content and utility of each scale are not elaborated herein.

Table 2.

Pain and emotion-related scale scores stratified by drug-refractory trigeminal neuralgia status and multi-stage feature selection workflow in the primary cohort

Characteristics	Initial Screening				Univariate Analysis			Multivariate Analysis		Selection Frequency (seeds; n = 20)
Characteristics	without DRTN (n = 31)	with DRTN (n = 104)	P value	B-H FDR	OR (95% CI)	P value		OR (95% CI)	P value	Selection Frequency (seeds; n = 20)
Pain and emotion-related scales (scores, mean ± SD)
VAS	6.37 ± 1.05	6.84 ± 0.90	0.012*	—	—	—		—	—	—
MPQ	16.32 ± 3.49	18.08 ± 2.48	0.007	0.026	1.23 (1.07–1.43)	0.005		1.18 (1.01–1.38)	0.036	13/20
PFPR	77.97 ± 30.73	87.88 ± 20.12	0.099							0/20
PSQ	71.45 ± 10.48	72.79 ± 11.25	0.543							0/20
PASS	40.45 ± 16.31	45.05 ± 15.55	0.170							0/20
PCS	23.68 ± 12.42	26.29 ± 9.42	0.286							0/20
PHQ-2	1.84 ± 0.90	2.14 ± 1.07	0.141	—	—	—	—		—	—
PHQ-9	7.00 ± 3.31	9.01 ± 4.28	0.015	0.043	1.15 (1.02–1.29)	0.021	1.10 (0.97–1.24)		0.152	4/20
GAD-7	7.55 ± 3.36	7.83 ± 2.76	0.499							0/20

Open in a new tab

Note: This table is descriptive only, presenting pain and emotion-related scale scores of the primary cohort and the feature selection workflow. All features adopted for modeling were derived within a nested cross-validation framework (20 random seeds): at each seed, the primary cohort underwent stratified splitting, after which feature selection was confined to the training cohort, followed by model evaluation on the withheld internal test cohort and validation cohort. Selection Frequency (seeds; n = 20) denotes the number of times a given feature was ultimately retained for modeling across the 20 training iterations. DRTN, drug-refractory trigeminal neuralgia; MPQ, McGill Pain Questionnaire; VAS, Visual Analogue Scale; PFPR, Penn Facial Pain Scale; PSQ, Pain Sensitivity Questionnaire; PASS, Pain Anxiety Symptoms Scale; PCS, Pain Catastrophizing Scale; PHQ-2, Patient Health Questionnaire-2; PHQ-9, Patient Health Questionnaire-9; GAD-7, Generalized Anxiety Disorder-7; B-H FDR, Benjamini–Hochberg False Discovery Rate; OR, odds ratio; CI, confidence interval. Continuous variables are presented as mean ± SD. P values for initial screening were calculated using Mann-Whitney U test or Independent t-test (Welch). Univariate and multivariate analyses were performed using binary logistic regression. Bold values indicate statistical significance (P < 0.05). Characteristic marked with * were excluded from further analyses due to severe multicollinearity. Specifically, VAS was removed as it is inherently embedded within the MPQ, resulting in substantial collinearity between the two scales

Imaging measurement parameters: All patients underwent high-resolution CT of the head and trigeminal nerve MRI using standardized scanning protocols. The evaluated imaging parameters included: neurovascular compression (NVC) grade [35] (grade 0: no compression; grade 1: vascular contact without indentation; grade 2: vascular contact with nerve indentation; grade 3: vascular contact causing nerve displacement, deformity, or atrophy), distance segment from NVC to the brainstem, offending vessel type, presence of brainstem compression, foramen ovale area, maximum diameter of the trigeminal porus, maximum area of Meckel’s cave, trigeminal pontine angle (TPA), petrous ridge angle (PRA), trigeminal nerve angle (TNA), and medial temporal lobe atrophy (MTA) grade, as detailed in Table 3. All imaging assessments were independently performed by two specialized physicians in a partially-blinded manner. Intraclass correlation coefficient (ICC) was used to evaluate inter-rater agreement; for continuous data, ICC > 0.75 was required and the mean value was adopted to minimize measurement error; for categorical and ordinal data, complete agreement between assessors was required for each patient, and discrepancies were adjudicated by a senior attending physician.

Table 3.

Imaging characteristics stratified by drug-refractory trigeminal neuralgia status and multi-stage feature selection workflow in the primary cohort

Characteristics	Initial Screening				Univariate Analysis		Multivariate Analysis		Selection Frequency (seeds; n = 20)
Characteristics	without DRTN (n = 31)	with DRTN (n = 104)	P value	B-H FDR	OR (95% CI)	P value	OR (95% CI)	P value	Selection Frequency (seeds; n = 20)
Relevant characteristics of the pontine segment of the TN
Neurovascular Compression (NVC)			< 0.001	< 0.001	3.078 (1.83–5.19)	< 0.001	3.454 (1.95–6.12)	< 0.001	20/20
0	7 (22.6)	14 (13.5)
1	22 (71.0)	25 (24.0)
2	2 (6.5)	31 (29.8)
3	0 (0.0)	34 (32.7)
Brainstem-side distance of NVC			0.498						0/20
Without	7 (22.6)	14 (13.5)
Proximal segment	12 (38.7)	48 (46.2)
Middle segment	6 (19.4)	27 (26.0)
Distal segment	6 (19.4)	15 (14.4)
Responsible Vessel			0.094						0/20
Without	7 (22.6)	14 (13.5)
SCA and AICA	23 (74.2)	72 (69.2)
Others	1 (3.2)	18 (17.3)
Brainstem Compression			0.085						0/20
Without	31 (100.0)	91 (87.5)
With	0 (0.0)	13 (12.5)
	Others
Foramen ovale area (cm²)	0.23 ± 0.03	0.2 ± 0.05	< 0.001	< 0.001	per 0.01 cm²: 0.94 (0.86–0.99)	0.028	per 0.01 cm²: 0.90 (0.83–0.98)	0.010	6/20
Maximum diameter of the trigeminal porus (cm)	0.5 ± 0.09	0.51 ± 0.08	0.282						0/20
Maximum area of Meckel’s cave (cm²)	0.49 ± 0.14	0.48 ± 0.13	0.931						0/20
Trigeminal pontine angle (TPA, °)	48.8 ± 10.69	50.66 ± 11.91	0.353						0/20
Petrous ridge angle (PRA, °)	82.47 ± 12.68	80.05 ± 13.69	0.372						0/20
Trigeminal nerve angle (TNA, °)	140.23 ± 11.62	143.78 ± 13.02	0.070						0/20
Medial Temporal Lobe Atrophy(MTA)			0.137						0/20
0	24 (77.4%)	57 (54.8%)
1	5 (16.1%)	19 (18.3%)
2	2 (6.5%)	18 (17.3%)
3	0 (0.0%)	8 (7.7%)
4	0 (0.0%)	2 (1.9%)

Open in a new tab

Note: This table is descriptive only, presenting imaging characteristics of the primary cohort and the feature selection workflow. All features adopted for modeling were derived within a nested cross-validation framework (20 random seeds): at each seed, the primary cohort underwent stratified splitting, after which feature selection was confined to the training cohort, followed by model evaluation on the withheld internal test cohort and validation cohort. Selection Frequency (seeds; n = 20) denotes the number of times a given feature was ultimately retained for modeling across the 20 training iterations. DRTN, drug-refractory trigeminal neuralgia; NVC, neurovascular compression; SCA, superior cerebellar artery; AICA, anterior Inferior Cerebellar Artery; TPA, trigeminal pontine angle; PRA, petrous ridge angle; TNA, trigeminal nerve angle; MTA, medial temporal lobe atrophy; B-H FDR, Benjamini–Hochberg False Discovery Rate; OR, odds ratio; CI, confidence interval. Continuous variables are presented as mean ± SD; categorical variables are presented as n (%). P values for initial screening were calculated using Mann-Whitney U test for continuous variables and Chi-square test for categorical variables. Univariate and multivariate analyses were performed using binary logistic regression. Bold values indicate statistical significance (P < 0.05). Red Inline graphic : To provide a clinically meaningful effect size, we rescaled the OR to a per-0.01 cm² increment

Feature selection

The feature selection process was conducted in two parts (Figs. 2 and 4).

Part 1 (Fig. 2): Feature selection for machine learning models was conducted within a nested cross-validation framework using 20 independent random seeds (70%/30% training/test split per seed; Fig. 2). Within each seed, feature selection was strictly confined to the training set to preclude data leakage. The process proceeded as follows: initial screening was performed through inter-group difference analysis for each of the three data dimensions; variables with severe multicollinearity were excluded; the Benjamini–Hochberg false discovery rate (FDR) correction was applied to control for multiplicity. Additionally, pain duration at the follow-up endpoint was excluded because it contains information on the duration of pharmacotherapy response, which is directly related to the study outcome, to prevent outcome data leakage. The remaining variables in each data domain were then subjected to univariate and multivariate binary logistic regression analyses within each training set, retaining statistically significant variables as candidate predictive features. The selection frequency of each feature—defined as the number of seeds (out of 20) in which it was retained for modeling—was recorded to quantify selection stability (Tables 1, 2 and 3).

Part 2 (Fig. 4): Univariate Cox proportional hazards regression was first performed on all candidate variables derived from the three data sources (baseline clinical characteristics, scale scores, and imaging features). To examine the robustness of variable selection, four univariate screening thresholds were applied in parallel: P < 0.05 (primary analysis) and P < 0.10, P < 0.15, and P < 0.20 (sensitivity analyses). Given that the MTA score is inherently age-associated, age was forcibly introduced into every multivariate model to adjust for potential confounding, regardless of its univariate significance.

Machine learning model construction and evaluation

This study employed five machine learning algorithms for DRTN binary classification modeling: support vector machine with radial basis function kernel (SVM-RBF), support vector machine with linear kernel (SVM-Linear), random forest (RF), extreme gradient boosting (XGBoost), and logistic regression (LR). The selection of these algorithms covers multiple modeling frameworks: linear models (LR, SVM-Linear), kernel-based models (SVM-RBF), and ensemble tree-based models (RF, XGBoost), thereby enabling a comprehensive comparison of different modeling approaches for the DRTN prediction task. The pathogenesis of DRTN involves the interactive effects of multi-dimensional factors, and both linear associations and nonlinear effects may exist between predictor variables and the outcome. A single modeling framework is insufficient to capture such complex data structures; by covering linear, kernel function, and ensemble learning paradigms, the suitability of different algorithms for the DRTN prediction task can be systematically evaluated.

The procedure was as follows: (1) All models were implemented using Python 3.9.7 (https://www.python.org), based on the Scikit-learn and XGBoost libraries; (2) explicit regularization and model structural constraints were applied according to the inherent characteristics of each algorithm to suppress overfitting from the perspective of model complexity, thereby enhancing model generalizability and suitability for real-world clinical applications; (3) grid search combined with 10-fold cross-validation was employed for hyperparameter optimization on the training cohort, and model performance was sequentially evaluated on the test cohort and the validation cohort; (4) to ensure the stability, reliability, and reproducibility of model performance evaluation, this study adopted a multi-random-seed strategy, using computer-generated sets of 20 different random seeds to independently perform training–test cohort partitioning and model parameter initialization. For each random seed, the complete analytical pipeline, including dataset partitioning, feature selection within the training set, model training, and performance evaluation, was executed, and the area under the receiver operating characteristic curve (AUC) was subsequently calculated. The mean AUC across the 20 seeds (mean ± standard deviation) served as the core evaluation metric. Traditional machine learning studies typically use a single random seed for dataset partitioning; however, the random split of training and test cohorts inherently introduces randomness that may cause fluctuations in model evaluation metrics. Particularly for this study, with a relatively limited sample size, the quality of a single data partition may lead to inflated or deflated model performance estimates, thereby affecting the credibility of the conclusions to some extent. The multi-random-seed strategy substantially mitigates the interference of random partitioning on model evaluation by repeatedly assessing model performance across multiple independent data partitioning schemes, yielding more robust and reproducible generalization performance estimates; (5) additionally, the median-seed model (i.e., the model trained using the random seed yielding the median AUC) was selected for subsequent calibration curve analysis, decision curve analysis (DCA), and SHAP feature interpretation, avoiding result bias from selecting a model with overly favorable or unfavorable performance. The above study design aligns with current recommendations for improving the reproducibility of medical machine learning research [36], and complies with the latest TRIPOD + AI guidelines for transparent reporting of prediction model studies [37].

Model performance was comprehensively quantified from multiple dimensions using the following metrics: AUC comprehensively reflects the overall discriminative capacity of the model across different thresholds; calibration curves with Brier Score and Hosmer–Lemeshow (H-L) P value were plotted to assess the agreement between predicted probabilities and observed outcomes; DCA was used to calculate the clinical net benefit of the model across different threshold probability intervals, evaluating the practical value of the model from a clinical decision-making perspective. Details are presented in Fig. 2.

SHapley Additive exPlanations (SHAP) analysis was employed to interpret the optimal machine learning model, elucidating the interpretable patterns of each feature’s contribution. The SHAP summary beeswarm plot visualized the direction and magnitude of each feature’s effect on model output. The feature importance bar chart based on SHAP values ranked features by mean absolute SHAP value, quantitatively characterizing the overall contribution of each feature to model prediction. Confusion matrices and predicted probability distribution histograms were plotted for both the test cohort and the temporally separated validation cohort, intuitively demonstrating the classification performance and prediction confidence distribution at the individual patient level. Details are presented in Fig. 3.

Fig. 3 — Interpretability and predictive performance of the optimal SVM-RBF model for drug-refractory trigeminal neuralgia. (A) SHAP summary beeswarm plot showing the impact and direction of each feature on model output. This figure quantifies the contribution magnitude and directionality of the five clinical features retained in the SVM-RBF model corresponding to the median-performing seed (by AUC) across multiple random splits, for predicting drug-refractory trigeminal neuralgia (DRTN) risk. Y-axis lists features ranked from highest to lowest overall predictive importance top-to-bottom; X-axis denotes SHAP values: positive values elevate DRTN risk while negative values reduce risk, with zero as the reference. Dot color encodes feature magnitude: blue for low feature values and red for high values. (B) SHAP-based feature importance bar chart ranking the five features by mean absolute SHAP value. X-axis denotes SHAP-based feature weight, with larger values indicating greater predictive contribution to DRTN; Y-axis lists five predictors sorted descendingly by importance from top to bottom. (C) Confusion matrix of the SVM-RBF model on the testing cohort. (D) Predicted probability distribution histogram for the testing cohort. (E) Confusion matrix on the temporally separated validation cohort. (F) Predicted probability distribution histogram for the temporally separated validation cohort. SHAP, SHapley Additive exPlanations; SVM, support vector machine; RBF, radial basis function; DRTN, drug-refractory trigeminal neuralgia

Cox regression model construction and evaluation

We defined time zero as the date when patients first received carbamazepine or oxcarbazepine therapy. This time point marks the commencement of standardized first-line medical treatment and serves as a well-defined, objective baseline for time-to-event analyses. The primary endpoint for Cox regression analysis was defined as the date of confirmed development of drug-resistant trigeminal neuralgia (DRTN), with diagnostic criteria specified in section “DRTN grouping criteria and detail”. For patients who maintained sustained therapeutic response and clinical stability on pharmacotherapy for more than two years without fulfilling the DRTN diagnostic thresholds, administrative censoring was applied at their last clinical follow-up. Additionally, subjects who remained free of DRTN onset by the study cutoff date were censored at their final follow-up time point.

Based on the four univariate Cox regression thresholds established in section “Feature selection”, Part 2, variables meeting each respective threshold were entered into a corresponding multivariate Cox regression model, with P < 0.05 as the multivariate inclusion criterion, yielding four parallel multivariate models (Fig. 4A and D). Given that the MTA score is inherently age-associated, age was forcibly introduced into every multivariate model to adjust for potential confounding, regardless of its univariate significance. The proportional hazards assumption was verified using Schoenfeld residual tests, with the non-proportionality P value used to confirm the assumption. Across all four screening thresholds, the MTA score remained a significant independent predictor after adjustment for age and all other retained variables (P < 0.05 in each model).

Using the primary-analysis model (univariate P < 0.05), the prognostic performance of an age-unadjusted and an age-adjusted multivariate Cox model was further compared. The discriminative capacity of the models was evaluated using the concordance index (C-index) and time-dependent area under the receiver operating characteristic curve (AUC) at 1, 3, and 5 years. The C-index provides an overall assessment of the model’s ability to distinguish between high-risk and low-risk individuals throughout the entire follow-up period, whereas time-dependent AUC characterizes the variation in discriminative capacity as a function of follow-up duration. Time-dependent AUC was calculated using the inverse probability of censoring weighting (IPCW) method to mitigate estimation bias introduced by censored observations. Kaplan–Meier survival curves were plotted for the high- and low-risk groups, and the log-rank test was used to assess the statistical significance of between-group survival differences, in both the training and validation cohorts (Fig. 4E and J). Calibration of the Cox regression models was additionally assessed.

Statistical analysis

Statistical analysis of clinical data was performed using Python 3.9.7 (https://www.python.org) and R (version 4.1.2, https://www.r-project.org/). For inter-group comparisons, normally distributed continuous variables were analyzed using independent-samples t-tests, non-normally distributed continuous variables using Mann–Whitney U tests, and categorical variables using chi-square tests or Fisher’s exact tests as appropriate. Binary logistic regression was employed for univariate and multivariate analyses of DRTN occurrence, with results presented as odds ratios (ORs) and 95% confidence intervals (CIs); Cox proportional hazards regression was used for time-to-event analysis, with results expressed as hazard ratios (HRs) and 95% CIs. Time-dependent AUC was calculated using the IPCW method to handle censored data. All statistical tests were two-sided, with P < 0.05 defined as statistically significant.

Results

Intergroup feature comparison and feature selection in the primary cohort

A total of 163 patients with primary trigeminal neuralgia were enrolled in this study, including 135 in the primary cohort (training cohort: n = 94; test cohort: n = 41) and 28 in the temporally separated validation cohort (Fig. 1). In the primary cohort, 104 patients were classified into the DRTN group and 31 into the non-DRTN group; in the temporally separated validation cohort, 20 patients were classified into the DRTN group and 8 into the non-DRTN group.

Tables 1, 2 and 3 are descriptive only, presenting the intergroup analyses and feature selection workflow for the primary cohort. All features adopted for modeling were derived within a nested cross-validation framework (20 random seeds). At each seed, the primary cohort underwent stratified splitting, after which feature selection was confined exclusively to the training cohort, followed by model evaluation on the withheld internal test cohort and the validation cohort.

Tables 1, 2 and 3 report the intergroup comparison results, including FDR-adjusted significance, univariate and multivariate logistic regression results, and cross-seed selection frequency for all candidate variables. Features selected through intergroup comparison are as follows:

(1) Clinical baseline characteristics (Table 1): pain involved extent (multivariate OR = 2.73, 95% CI: 1.20–6.20, P = 0.016, selection frequency: 11/20 seeds), concomitant continuous pain (multivariate OR = 9.40, 95% CI: 1.14–77.77, P = 0.038, selection frequency: 7/20 seeds), BMI (multivariate OR = 1.19, 95% CI: 1.01–1.39, P = 0.034, selection frequency: 4/20 seeds), age at first pain onset (multivariate OR = 1.07, 95% CI: 1.01–1.12, P = 0.020, selection frequency: 3/20 seeds). The difference in hypertension risk stratification was statistically significant in univariate analysis (univariate OR = 1.70, 95% CI: 1.16–2.47, P = 0.006), whereas no statistical significance was observed in multivariate analysis (multivariate OR = 1.26, 95% CI: 0.83–1.91, P = 0.270); nevertheless, its selection frequency reached 9 out of 20 random seeds. Furthermore, sex, pain laterality, and immune-related diseases did not differ significantly between groups, and each of these variables was selected in only 1 of the 20 random seeds. Height, weight, affected trigeminal branches, and highest systolic blood pressure were excluded during initial screening due to multicollinearity; the remaining variables showed no statistically significant differences (P > 0.05).

(2) Pain and emotion-related scale scores (Table 2): MPQ (multivariate OR = 1.18, 95% CI: 1.01–1.38, P = 0.036, selection frequency: 13/20 seeds). PHQ-9 score was statistically significant in univariate analysis (univariate OR = 1.15, 95% CI: 1.02–1.29, P = 0.043) but not in multivariate analysis (multivariate OR = 1.10, 95% CI: 0.97–1.24, P = 0.152); nevertheless, its selection frequency reached 4 out of 20 random seeds. VAS scores were excluded due to severe multicollinearity. The remaining scales (PFPR, PSQ, PASS, PCS, PHQ-2, GAD-7) showed no statistically significant inter-group differences (P > 0.05).

(3) Imaging characteristics (Table 3): NVC grade (multivariate OR = 3.454, 95% CI: 1.95–6.12, P < 0.001, selection frequency: 20/20 seeds) and foramen ovale area (per 0.01 cm²; multivariate OR = 0.90, 95% CI: 0.83–0.98, P = 0.010, selection frequency: 6/20 seeds). The remaining imaging indicators (NVC-to-brainstem distance segment, offending vessel type, brainstem compression, maximum diameter of the trigeminal porus, maximum area of Meckel’s cave, TPA, PRA, TNA, and MTA grade) showed no statistically significant intergroup differences (P > 0.05).

An exploratory subgroup analysis was performed to compare clinical characteristics, pain and emotion-related scale scores, and imaging characteristics between the analgesic-failure and analgesic-intolerance subgroups within the DRTN cohort. No statistically significant differences (P > 0.05) were observed between subgroups in either the primary or validation cohort across these three domains (Supplementary Tables 1–3).

Machine learning model performance

The performance of the five machine learning models in the training, test, and validation cohorts is summarized in Table 4; Fig. 2. All AUC values are presented as mean ± standard deviation across 20 random seeds. Among all models, the SVM-RBF model demonstrated comparatively favorable and stable overall predictive performance, with average AUCs of 0.9269 ± 0.0412 in the training cohort, 0.8235 ± 0.0490 in the test cohort, and 0.8063 ± 0.0577 in the validation cohort. The RF model achieved a comparable average AUC of 0.8075 ± 0.0422 in the validation cohort, followed by XGBoost (0.7941 ± 0.0423), linear SVM (0.7775 ± 0.0593), and logistic regression (0.7741 ± 0.0459).

Table 4.

Performance comparison of machine learning models (Average AUC, Mean ± SD)

Models	Training Cohorts (n = 94)	Testing Cohorts (n = 41)	Temporally separated validation Cohorts (n = 28)
SVM-RBF	0.9269 ± 0.0412	0.8235 ± 0.0490	0.8063 ± 0.0577
SVM-Linear	0.8676 ± 0.0378	0.8077 ± 0.0683	0.7775 ± 0.0593
Random Forest	0.9532 ± 0.0224	0.8658 ± 0.0527	0.8075 ± 0.0422
XGBoost	0.9308 ± 0.0268	0.8567 ± 0.0503	0.7941 ± 0.0423
Logistic Regression	0.8730 ± 0.0352	0.8301 ± 0.0578	0.7741 ± 0.0459

Open in a new tab

Note: AUC, Area Under the Receiver Operating Characteristic Curve; SVM, Support Vector Machine; RBF, Radial Basis Function; Average AUC was calculated across multiple random seeds to enhance model stability, reliability and reproducibility, and minimize random bias

Calibration curves and decision curve analysis, generated based on median performance metrics across the 20 random seed replicates (Fig. 2), showed that the SVM-RBF model yielded non-significant H–L test P values in all three cohorts: training (Brier score = 0.073, H–L P = 0.250), test (Brier score = 0.087, H–L P = 0.550), and validation (Brier score = 0.124, H–L P = 0.354). Of note, only the RF model exhibited statistically significant miscalibration in both the test and validation cohorts (H–L P < 0.05), indicating systematic deviation between predicted and observed risks upon generalization to other samples. However, the H–L test has inherently limited statistical power in small samples (n = 28; n = 41); thus, non-significant P values should be interpreted with caution. Decision curve analysis further demonstrated that the SVM-RBF model achieved favorable clinical net benefit across a wide range of threshold probabilities in the validation cohort (Fig. 2).

SHAP analysis of the optimal model (SVM-RBF)

SHAP analysis was performed on the SVM-RBF model corresponding to the median-performing seed across the 20 random splits. In this seed, five features contributed to the model: NVC grade, MPQ, concomitant continuous pain, pain involved extent, and hypertension risk stratification; foramen ovale area, BMI, and age at first pain onset were not retained, consistent with their lower or variable selection frequencies across seeds (Tables 1, 2 and 3). The SHAP summary beeswarm plot (Fig. 3A) shows the impact and direction of each feature on model output, with features ranked from highest to lowest overall importance (top to bottom). SHAP analysis also revealed a nonlinear effect of NVC grade: when NVC grade reached 3, SHAP values increased substantially.

The feature importance ranking based on mean absolute SHAP values (Fig. 3B) was consistent with the beeswarm plot: NVC grade contributed the highest SHAP weight, followed by MPQ, concomitant continuous pain, pain involved extent, and hypertension risk stratification. Aggregated by data dimension, the imaging feature (NVC grade) contributed a mean SHAP value of 0.1691, the single scale score feature (MPQ) contributed 0.0796, and the three clinical features (pain involved extent, concomitant continuous pain, and hypertension risk stratification) collectively contributed a mean SHAP value of 0.2055.

The test cohort confusion matrix (Fig. 3C) showed: true negatives = 5, true positives = 30, false positives = 4, and false negatives = 2. The predicted probability distribution histogram (Fig. 3D) demonstrated that DRTN cases were predominantly distributed in the high-probability range (0.7–1.0). The temporally separated validation cohort confusion matrix (Fig. 3E) showed: true negatives = 3, true positives = 19, false positives = 5, and false negatives = 1. The predicted probability distribution histogram (Fig. 3F) revealed that DRTN cases were predominantly distributed in the high-probability range (0.8-1.0).

Cox regression models

Univariate and multivariate regression analysis

Univariate Cox proportional hazards regression using the primary screening threshold (P < 0.05) identified four predictors significantly linked to time of DRTN onset: concomitant continuous pain (HR = 1.584, 95% CI: 1.036–2.420, P = 0.034), pain involved extent (HR = 1.531, 95% CI: 1.079–2.172, P = 0.017), body mass index (BMI; HR = 1.072, 95% CI: 1.010–1.137, P = 0.022), and medial temporal lobe atrophy (MTA) score (HR = 1.234, 95% CI: 1.034–1.471, P = 0.020). Age exhibited a borderline non-significant univariate association (HR = 1.020, 95% CI: 1.000–1.040, P = 0.051; Table 5).

Table 5.

Univariate and multivariate Cox regression analysis of characteristics associated with prognosis under primary analysis

Characteristics	Univariate Analysis				Multivariate Analysis (Age-adjusted)
Characteristics	HR (95% CI)	P	HR (95% CI)	P	HR (95% CI)	P
Concomitant Continuous Pain	1.584 (1.036–2.420)	0.034	1.406 (0.910–2.173)	0.125	1.370 (0.888–2.114)	0.154
Pain Involved Extent	1.531 (1.079–2.172)	0.017	1.539 (1.072–2.211)	0.020	1.608 (1.109–2.333)	0.012
Body Mass Index	1.072 (1.010–1.137)	0.022	1.054 (0.991–1.121)	0.094	1.052 (0.989–1.119)	0.111
MTA	1.234 (1.034–1.471)	0.020	1.257 (1.046–1.510)	0.015	1.261 (1.048–1.518)	0.014
Age	1.020 (1.000-1.040)	0.051	—	—	1.020 (1.000–1.040)	0.056

Open in a new tab

Note: HR, hazard ratio; CI, confidence interval; MTA, medial temporal lobe atrophy. Characteristics with P < 0.05 in univariate Cox regression were included in the multivariate Cox regression model. Bold values indicate statistical significance (P < 0.05)

The four significant univariate variables were entered into two multivariate Cox models (age-unadjusted and age-adjusted, with age forced into the second model to mitigate age-related confounding of MTA). In the age-unadjusted model, only pain involved extent (HR = 1.539, 95% CI: 1.072–2.211, P = 0.020) and MTA score (HR = 1.257, 95% CI: 1.046–1.510, P = 0.015) remained independent predictors; concomitant continuous pain and BMI lost statistical significance, showing their crude univariate associations were mediated by the two retained markers. After age adjustment, pain involved extent (HR = 1.608, 95% CI: 1.109–2.333, P = 0.012) and MTA score (HR = 1.261, 95% CI: 1.048–1.518, P = 0.014) stayed significant, while age itself remained non-significant (HR = 1.020, 95% CI: 1.000–1.040, P = 0.056) (Tables 5 and 6; Fig. 4).

Table 6.

Predictive performance comparison between age-unadjusted and age-adjusted multivariate Cox regression models in temporally separated validation cohorts

Metric	Cox Regression Model
Metric	Multivariate (characteristics, n = 4)	Multivariate (characteristics, n = 4 + Age)
Primary Cohorts ( n = 135)
C-index	0.644	0.648
1-year AUC (95% CI)	0.740 (0.639–0.845)	0.734 (0.623–0.835)
3-year AUC (95% CI)	0.644 (0.550–0.731)	0.672 (0.581–0.757)
5-year AUC (95% CI)	0.671 (0.594–0.753)	0.677 (0.592–0.756)
High risk, n	68	68
Low risk, n	67	67
Temporally Separated Validation Cohorts ( n = 28)
C-index	0.619	0.626
1-year AUC (95% CI)	0.693 (0.493–0.875)	0.700 (0.497–0.885)
3-year AUC (95% CI)	0.607 (0.417–0.797)	0.629 (0.433–0.823)
5-year AUC (95% CI)	0.572 (0.331–0.810)	0.606 (0.375–0.813)
High risk, n	17	17
Low risk, n	11	11

Open in a new tab

Note: AUC, Area Under the Receiver Operating Characteristic Curve; CI, confidence interval. The univariate Cox regression model includes individual significant predictors separately; the multivariate Cox regression model includes all significant predictors simultaneously. Risk groups were stratified by the median risk score (cutoff value) in the training cohort

Sensitivity analyses with three relaxed univariate screening thresholds (P < 0.10, P < 0.15, P < 0.20) generated additional multivariate models (Fig. 4A and D). Pain involved extent and MTA score maintained consistent independent significance (P < 0.05) across all four threshold-based models after mutual and age adjustment, whereas other covariates lacked stable predictive value. Schoenfeld residual tests validated the proportional hazards assumption for all retained variables.

Cox model performance and time-dependent discrimination analysis

Based on the primary multivariate model (P < 0.05 univariate threshold), we compared the performance of age-unadjusted and age-adjusted Cox models in primary (n = 135) and validation (n = 28) cohorts (Table 6; Fig. 4E and J). In the primary cohort, the unadjusted and age-adjusted models yielded highly consistent C-indexes (0.644 vs. 0.648) and comparable 1-, 3-, and 5-year time-dependent AUCs, with marginal numerical differences observed between the two models (Table 6; Fig. 4F and I). Similarly, the validation cohort presented similar C-indexes (0.619 vs. 0.626) and temporal AUC profiles for unadjusted and age-adjusted models (Table 6; Fig. 3G and J). Overall, age adjustment produced negligible changes in model performance (maximal C-index difference = 0.007, maximal AUC difference = 0.034).

Furthermore, both models demonstrated moderate discriminative ability in the short to medium term (1 and 3 years), with reduced predictive performance in the long term (5 years). Kaplan–Meier and log-rank analyses confirmed significant DRTN-free survival differences between risk strata for both age-unadjusted (Fig. 4E) and age-adjusted (Fig. 4H) models, verifying the reliable risk-stratification capability of the established Cox models despite their moderate absolute discriminative power.

Discussion

This study systematically collected and integrated multi-source heterogeneous data encompassing clinical baseline characteristics, pain and emotional scale scores, and imaging measurement parameters. Leveraging machine learning algorithms and Cox regression models, we constructed and preliminarily validated an exploratory model system for predicting DRTN onset risk and onset timing, aiming to deepen the understanding of TN disease progression and provide preliminary reference for future prospective research and large-sample multicenter studies.

Multi-domain feature interpretation

The integration of heterogeneous data from multiple dimensions represents an important methodological strength of this study. In clinical practice, we frequently encounter the following dilemma: patients’ actual pain experience does not fully correspond to the severity of ipsilateral neurovascular compression (NVC), and moderate-to-severe NVC signs can also be observed in some healthy individuals. Furthermore, during regular follow-up, we observed that some patients with comparable clinical baseline profiles exhibited substantial heterogeneity in their pharmacotherapy and surgical outcomes. Although NVC is a key initiating factor in TN [1], we hypothesize that the specific disease progression and clinical manifestations are the result of multifactorial interactions. Therefore, we sought to utilize the most comprehensive multi-dimensional data available to dissect the heterogeneity of patients’ clinical symptoms and pharmacotherapy prognosis from the perspective of phenotypic characteristics.

Analysis of intergroup baseline clinical characteristics within the primary cohort identified pain involved extent, concomitant continuous pain, BMI, and age at first pain onset as independent predictors in multivariate logistic regression. Additionally, hypertension risk stratification, though not statistically significant in multivariate analysis of the primary cohort, was retained with high selection frequency (9/20 seeds) across the nested CV framework, reflecting its potential predictive effects in partial training subsets. Pain involved extent was the most stably selected clinical feature (11/20 seeds), suggesting that a broader spatial distribution of pain corresponds to more extensive nerve involvement and more severe pathological damage, with a correspondingly higher probability of pharmacotherapy failure. Concomitant continuous pain (7/20 seeds) was positively associated with DRTN risk in both univariate and multivariate analyses; however, given that only 1 of the 31 non-DRTN patients had concomitant continuous pain, the data structure approached near-complete separation, which yields unstable OR estimates with wide confidence intervals (multivariate OR = 9.40, 95% CI: 1.14–77.77). Nevertheless, the direction of this association remained consistent in SHAP analyses and retains a certain degree of reasonableness. Di Stefano et al. demonstrated through neuroimaging evidence that TN patients with concomitant continuous pain exhibit trigeminal nerve root atrophy, suggesting that continuous pain is not merely a superimposed clinical symptom but rather reflects more severe structural nerve damage and central sensitization processes [22]. Additionally, age at first pain onset (3/20 seeds) and BMI (4/20 seeds) were also independent predictive factors; we speculate that older age at onset may be associated with diminished repair capacity due to neurodegeneration [38], while elevated BMI may influence treatment response through systemic inflammatory states and alterations in pharmacokinetics [39, 40].

In the analysis of pain and emotion-related scale scores between groups within the primary cohort, MPQ score was the sole retained predictor (13/20 seeds). The MPQ quantifies pain experience across sensory, affective, and evaluative dimensions [41]; higher MPQ scores typically indicate more complex pain qualities and more pronounced central sensitization. Although the PHQ-9 (4/20 seeds) score was statistically significant in univariate analysis, it did not reach significance in multivariate analysis, suggesting that the predictive information contained in depressive symptoms may be substantially captured by the MPQ. This pattern aligns with evidence of a close interactive association between pain and depression in TN patients [24].

Analysis of imaging parameters revealed that NVC grade was the most influential predictor, being retained in all 20 seeds and ranked highest in SHAP feature importance. This finding is consistent with the well-established pathophysiological role of NVC in TN [1–3]. Maarbjerg et al. prospectively demonstrated that severe NVC causing nerve displacement or atrophy is strongly associated with the symptomatic side in classical TN [35]. The SHAP analysis in this study further revealed a nonlinear effect of NVC grade on model prediction: when NVC grade reached 3 (vascular compression causing nerve displacement/deformity/atrophy), SHAP values increased sharply, suggesting that severe compression may more readily facilitate the transition from drug responsiveness to drug resistance in TN patients, potentially reflecting that higher NVC grades often entail more severe structural nerve damage, thereby increasing the difficulty of achieving analgesia through sodium channel blocker-mediated modulation of neuronal excitability. Additionally, foramen ovale area was identified as a significant imaging predictor of DRTN. To improve clinical interpretability, the effect was rescaled to per 0.01 cm² increments. After multivariate adjustment, the OR was 0.90 (95% CI: 0.83–0.98, P = 0.010), indicating that each 0.01 cm² increase in foramen ovale area was associated with an approximately 10% reduction in the odds of DRTN. From an anatomical perspective, foramen ovale stenosis can lead to entrapment of the V3 branch during its transcranial course, resulting in chronic mechanical nerve stimulation and peripheral sensitization. Li et al. confirmed through morphometric analysis that the foramen ovale area in patients with primary trigeminal neuralgia is significantly smaller than in healthy controls, and that a narrowed foramen ovale may contribute to TN pathogenesis through persistent compression of V3 [20]. Subsequent research further identified a significant association between foramen ovale stenosis and concomitant continuous pain in TN [42], suggesting that bony entrapment and vascular compression may synergistically exacerbate nerve damage. Notably, although MVD can relieve vascular compression, it cannot alleviate bony entrapment caused by foramen ovale stenosis, which may partially explain the suboptimal postoperative outcomes in patients with concomitant foramen ovale stenosis.

Machine learning model performance analysis

Among the five candidate algorithms, RF achieved the numerically highest average AUCs across all three cohorts: 0.953 ± 0.022 (training), 0.866 ± 0.053 (test), and 0.808 ± 0.042 (validation) (Table 4). However, RF exhibited statistically significant miscalibration in both the test and validation cohorts (H-L P < 0.05), indicating systematic deviation between predicted probabilities and observed event rates. This reflects an inherent limitation of tree-based ensemble methods: the hard partitioning of decision trees drives ensemble-averaged probability estimates toward extremes, producing overconfident predictions that are poorly calibrated. As another tree-based ensemble algorithm, XGBoost also exhibited a similar but milder issue, indirectly suggesting that, given the current sample size, this limitation may be inherent to tree-based ensemble modeling.

Overall, SVM-RBF offered the most acceptable balance of discrimination and calibration among the evaluated algorithms under the present study conditions. Its validation AUC (0.806 ± 0.058) was essentially identical to that of RF (0.808 ± 0.042), and the H-L test did not reject calibration in any cohort (training P = 0.250, test P = 0.550, validation P = 0.354). Decision curve analysis indicated that SVM-RBF yielded favorable net clinical benefit across a broad range of threshold probabilities in the validation cohort. This advantage stems from the capacity of the RBF kernel function to map input features into a high-dimensional Hilbert space, thereby constructing nonlinear decision boundaries in the feature space that capture complex nonlinear associations between predictor variables and the outcome. Furthermore, the discriminative advantage of the three nonlinear models (SVM-RBF, RF, XGBoost) over the two linear approaches, linear SVM (validation AUC = 0.778 ± 0.059) and logistic regression (validation AUC = 0.774 ± 0.046), indirectly corroborates that DRTN pathogenesis involves interactive effects across multi-dimensional factors.

In the validation cohort (n = 28), the SVM-RBF model at the median-performing seed correctly identified 19 of 20 DRTN cases (sensitivity = 95.0%, exact binomial 95% CI: 75.1%–99.9%). Although the model exhibited relatively high sensitivity in the present study, these performance estimates are subject to considerable uncertainty and should be interpreted as exploratory owing to the small validation sample size (n = 28) and wide confidence intervals. Moreover, 5 non-DRTN patients were misclassified as DRTN, indicating that model specificity still needs optimization and requires further evaluation in larger samples.

Cox model performance analysis

This study employed Cox regression modeling to analyze the temporal dynamics of DRTN onset, providing complementary prognostic information. Multivariate Cox regression, corroborated by a four-threshold sensitivity analysis, identified pain involved extent and MTA score as potentially independent prognostic markers of the timing of DRTN.

In the age-adjusted multivariate model, each additional affected trigeminal nerve branch was associated with a 61% increase in the hazard of DRTN onset (HR = 1.608, 95% CI: 1.109–2.333, P = 0.012). From a pathophysiological perspective, there is a plausible intrinsic association between broader pain involved and earlier DRTN onset: more extensive trigeminal nerve involvement indicates more severe nerve damage, which may accelerate the progression of TN.

The MTA score was independently associated with DRTN timing after forced adjustment for age (HR = 1.261, 95% CI: 1.048–1.518, P = 0.014), whereas age itself did not reach significance (HR = 1.020, 95% CI: 1.000–1.040, P = 0.056). In recent years, multiple TN-related neuroimaging studies have consistently demonstrated that the medial temporal lobe region—centered on the hippocampus and amygdala—is deeply involved in the affective-cognitive dimension of pain processing in TN and partially participates in the modulation of descending pain pathways [24, 26, 43–52].Medial temporal atrophy represents structural degeneration of the medial temporal brain tissue, and previous studies have confirmed its association with pain chronification, serving as a characteristic marker of adverse central neuroplastic changes [26, 53]. Therefore, we hypothesize that medial temporal lobe atrophy may compromise the central regulatory capacity for TN through multiple mechanisms, thereby accelerating the progression to DRTN.

Although the Cox models in this study demonstrated moderate discriminative capacity at the 1-year and 3-year follow-up time points, their discriminative efficacy declined when follow-up extended to 5 years. This finding is largely consistent with our expectations. The potential reasons for the time-dependent attenuation of model discriminative capacity include: (1) the multivariable Cox regression models incorporated merely 4 to 5 variables, a relatively small feature set that failed to fully capture the multi-dimensional pathogenic factors associated with long-term treatment outcomes; (2) as the follow-up period lengthens, various time-varying confounders that cannot be captured by baseline data accumulate progressively, gradually attenuating the predictive efficacy of baseline covariates. Medication adherence adjustments, dosage modifications, newly developed comorbidities, and pain phenotype evolution over the disease course were not incorporated into baseline data but may influence the trajectory of long-term pharmacotherapy response. The Cox models, relying solely on baseline indicators for modeling, cannot incorporate such dynamic changes; (3) the number of remaining subjects who developed DRTN during the later follow-up period progressively decreased, reducing the effective sample size for estimating time-dependent discriminative capacity and precluding precise determination of the model’s long-term prognostic value. Although the Cox models demonstrated limited long-term predictive capacity, the log-rank test indicated that they could preliminarily stratify patients by risk, suggesting potential utility in short-to-medium-term prognostic assessment that warrants further validation.

Limitation

This study has the following limitations that should be considered when interpreting the conclusions: (1) Prevalence enrichment and calibration concern: The DRTN prevalence in our modeling cohort was 77% (104/135), higher than the 30–50% refractoriness rate reported in real-world populations. This prevalence enrichment is primarily driven by poor follow-up adherence among patients with well-controlled pain, who were disproportionately excluded and thereby introduced bias in case distribution. Therefore, model predicted probabilities are calibrated only on this enriched high-prevalence cohort, with potential calibration drift observed upon extrapolation to routine clinical practice. Furthermore, the H-L test has limited statistical power in small sample sizes, compounding the aforementioned limitation; (2) The study design cannot completely eliminate selection bias and information bias. A total of 37 subjects across both cohorts were lost to follow-up, and the systematic exclusion of these cases may introduce attrition bias. Additionally, during follow-up, we observed that patients with satisfactory pain control and relatively stable disease were less motivated to attend follow-up visits and reluctant to undergo further examinations and medication guidance, whereas those with intractable pain were more inclined to continue seeking medical care, resulting in category imbalance in the follow-up population, which is the primary reason for the unbalanced sample sizes between groups in this study; (3) This was a single-center exploratory study with a limited sample size, and all cases were sourced from a single institution. As such, although the validation and modeling cohorts were separated by enrollment period, systematic bias between the two cannot be fully excluded. Consequently, while these findings carry preliminary reference value, they should be interpreted with caution. Future large-sample, multicenter cohort studies are warranted to more precisely evaluate model performance across different follow-up intervals and to validate the generalizability of these results; (4) The existing measurement precision and feature selection strategy have inherent limitations. Although the variables included in this study are relatively comprehensive in dimension, all selected features are conventional phenotypic indicators; particularly regarding imaging parameters, conventional imaging indicators cannot capture tissue microstructural and functional information compared with radiomics, deep learning, voxel-based morphometry (VBM), source-based morphometry (SBM), and functional MRI (fMRI) and other high-precision methods and techniques. Furthermore, this study adopted relatively stringent screening methods, which may have excluded variables with potential predictive value that did not individually reach statistical significance. In summary, we retrospectively reviewed and analyzed the data from cases that the model failed to correctly classify or predict, and identified some interesting associations. Although these associations did not reach statistical significance at the group level, we believe they may provide informative insights for future research, as detailed in Fig. 5.

Fig. 5 — Imaging manifestations of selected cases incorrectly classified and predicted by the model. All three patients had low-grade NVC (grade 0–1) yet progressed to drug-refractory trigeminal neuralgia (DRTN) within a short period. (A, B) A 47-year-old woman with right-sided trigeminal neuralgia who experienced early recurrence following sequential microvascular decompression (MVD) and percutaneous balloon compression (PBC): axial images show that the right trigeminal nerve (red arrow) is markedly attenuated compared with the contralateral side (blue arrow); sagittal images reveal a sharpened petrous ridge with significant stenosis of the trigeminal porus. (C, D) A 66-year-old man with left-sided trigeminal neuralgia who had suboptimal outcomes after surgical treatment at an outside institution: axial and coronal images demonstrate a tortuous vertebrobasilar artery compressing the adjacent pontomedullary region (red arrows). (E, F) A 53-year-old woman with right-sided trigeminal neuralgia who achieved limited benefit from MVD and continued to require pharmacological pain control postoperatively: axial and sagittal images show a small bony protrusion within Meckel’s cave compressing the adjacent trigeminal ganglion (red arrows). These cases suggest that, beyond classic neurovascular compression, other regional structural abnormalities—such as bony stenosis/anomalies and brainstem compression—may compromise the classification accuracy of the model; furthermore, there may exist abnormalities that are not visually discernible, which provides valuable implications for future research

Finally, as discussed in section “Cox model performance analysis”, we have established that Cox models relying solely on baseline indicators are insufficient to meet the clinical need for long-term prognostic assessment of DRTN. Therefore, in future studies, we will consider leveraging the aforementioned high-precision methods and techniques to address the limited and time-attenuating discriminative capacity of Cox regression models.

Conclusion

This study integrated multi-source heterogeneous data encompassing clinical baseline characteristics, pain and emotional scale scores, and imaging measurement parameters to construct and validate prediction models for DRTN onset risk and onset timing. The SVM-RBF model based on the radial basis function kernel demonstrated the best overall performance in DRTN risk prediction, and SHAP analysis revealed NVC grade as a key driving feature. Cox regression analysis identified pain involved extent and MTA score as significant, independent candidate prognostic markers for DRTN onset timing, though the time-dependent discriminative capacity of the models was limited. These findings provide preliminary evidence for early risk stratification and individualized treatment planning in TN patients receiving pharmacotherapy; however, given the limited long-term predictive capability of the Cox models and the single-center exploratory design, these results should be interpreted with caution. Future studies with large-sample, multi-center prospective cohorts are needed to further validate these conclusions and to explore superior survival time prediction modeling strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1^{(35.8KB, docx)}

Supplementary Material 2^{(271KB, docx)}

Supplementary Material 3^{(31.6KB, docx)}

Abbreviations

TN: Trigeminal neuralgia
MVD: Microvascular decompression
DRTN: Drug-refractory trigeminal neuralgia
ML: Machine learning
ICHD-3: International Classification of Headache Disorders 3rd Edition
MRI: Magnetic resonance imaging
CT: Computed tomography
BNI: Barrow Neurological Institute
PACS: Picture archiving and communication system
CTN: Classical trigeminal neuralgia
ITN: Idiopathic trigeminal neuralgia
BMI: Body mass index
OR: Odds ratio
VAS: Visual analog scale
MPQ: McGill Pain Questionnaire
PFPR: Penn Facial Pain Scale
PSQ: Pain Sensitivity Questionnaire
PASS: Pain Anxiety Symptoms Scale
PCS: Pain Catastrophizing Scale
PHQ-2: Patient Health Questionnaire 2-item
PHQ-9: Patient Health Questionnaire 9-item
GAD-7: Generalized Anxiety Disorder 7-item scale
NVC: Neurovascular compression
MTA: Medial temporal lobe atrophy
TPA: Trigeminal pontine angle
PRA: Petrous ridge angle
TNA: Trigeminal nerve angle
ICC: Intraclass correlation coefficient
FDR: False discovery rate
LR: Logistic regression
RF: Random forest
XGBoost: Extreme gradient boosting
SVM-Linear: Support vector machine with linear kernel
SVM-RBF: Support vector machine with radial basis function kernel
SVM: Support vector machine
RBF: Radial basis function
AUC: Area under the curve
SHAP: SHapley Additive exPlanations
DCA: Decision curve analysis
C-index: Concordance index
IPCW: Inverse probability of censoring weighting
CI: Confidence interval
HR: Hazard ratio
fMRI: functional magnetic resonance imaging
VBM: Voxel-based morphometry
SBM: Source-based morphometry
B-H FDR: Benjamini–Hochberg false discovery rate
SD: Standard deviation
ROC: Receiver operating characteristic

Author contributions

Hongan Yang was responsible for drafting the original manuscript. Baijintao Sun performed raw data extraction and collection of all follow-up data. Mingtao Li and Xuezhao Yang independently extracted and organized imaging data, respectively. Mingyan Li was responsible for figure and table construction. The conception and overall framework design of this research thesis were completed by Chuan Zhang. Hanfeng Yang participated in manuscript revision, final verification of all data, language polishing and editing of the full text.

Funding

Special Project for the Central Government to Guide the Development of Local Science and Technology in Sichuan Province (2024ZYD0092). Youth Science and Technology Talent Program of Nanchong Science and Technology Bureau (25KJCXRC0012). Research and Development Plan Project of Affiliated Hospital of North Sichuan Medical College (2023PTZK002, 2024GC012 and 2024MPZK008). Shanghai Yangtze River Delta Business Innovation and Technology Foundation Project (SC-KYF-202503002).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Hongan Yang and Baijintao Sun contributed equally to this work and share first authorship.

Contributor Information

Chuan Zhang, Email: cbyzc@qq.com.

Hanfeng Yang, Email: yhfctjr@163.com.

References

1.Ashina S, Robertson CE, Srikiatkhachorn A et al (2024) Trigeminal neuralgia. Nat Rev Dis Primers 10:39. 10.1038/s41572-024-00523-z [DOI] [PubMed] [Google Scholar]
2.Cruccu G, Di Stefano G, Truini A (2020) Trigeminal Neuralgia. N Engl J Med 383:754–762. 10.1056/NEJMra1914484 [DOI] [PubMed] [Google Scholar]
3.Bendtsen L, Zakrzewska JM, Heinskou TB et al (2020) Advances in diagnosis, classification, pathophysiology, and management of trigeminal neuralgia. Lancet Neurol 19:784–796. 10.1016/S1474-4422(20)30233-7 [DOI] [PubMed] [Google Scholar]
4.Tölle T, Dukes E, Sadosky A (2006) Patient burden of trigeminal neuralgia: results from a cross-sectional survey of health state impairment and treatment patterns in six European countries. Pain Pract 6:153–160. 10.1111/j.1533-2500.2006.00079.x [DOI] [PubMed] [Google Scholar]
5.Zakrzewska JM, Wu J, Mon-Williams M et al (2017) Evaluating the impact of trigeminal neuralgia. Pain 158:1166–1174. 10.1097/j.pain.0000000000000853 [DOI] [PubMed] [Google Scholar]
6.Spatz AL, Zakrzewska JM, Kay EJ (2007) Decision analysis of medical and surgical treatments for trigeminal neuralgia: how patient evaluations of benefits and risks affect the utility of treatment decisions. Pain 131:302–310. 10.1016/j.pain.2007.02.009 [DOI] [PubMed] [Google Scholar]
7.Bendtsen L, Zakrzewska JM, Abbott J et al (2019) European Academy of Neurology guideline on trigeminal neuralgia. Eur J Neurol 26:831–849. 10.1111/ene.13950 [DOI] [PubMed] [Google Scholar]
8.Rana MH, Khan AAG, Khalid I et al (2023) Therapeutic Approach for Trigeminal Neuralgia: A Systematic Review. Biomedicines 11:2606. 10.3390/biomedicines11102606 [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Pergolizzi JV, LeQuang JA, El-Tallawy SN et al (2024) An update on pharmacotherapy for trigeminal neuralgia. Expert Rev Neurother 24:773–786. 10.1080/14737175.2024.2365946 [DOI] [PubMed] [Google Scholar]
10.Gronseth G, Cruccu G, Alksne J et al (2008) Practice parameter: the diagnostic evaluation and treatment of trigeminal neuralgia (an evidence-based review): report of the Quality Standards Subcommittee of the American Academy of Neurology and the European Federation of Neurological Societies. Neurology 71:1183–1190. 10.1212/01.wnl.0000326598.83183.04 [DOI] [PubMed] [Google Scholar]
11.Li S, Chen H, Chen J et al (2024) Predicting long-term outcomes in patients with classical trigeminal neuralgia following microvascular decompression with an MRI-based radiomics nomogram: a multicentre study. Eur Radiol. 10.1007/s00330-024-10775-8 [DOI] [PubMed] [Google Scholar]
12.Andersen ASS, Heinskou TB, Rochat P et al (2022) Microvascular decompression in trigeminal neuralgia - a prospective study of 115 patients. J Headache Pain 23:145. 10.1186/s10194-022-01520-x [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Heinskou TB, Rochat P, Maarbjerg S et al (2019) Prognostic factors for outcome of microvascular decompression in trigeminal neuralgia: A prospective systematic study using independent assessors. Cephalalgia 39:197–208. 10.1177/0333102418783294 [DOI] [PubMed] [Google Scholar]
14.Xie ME, Halbert-Elliott K, Nair SK et al (2024) External assessment of preoperative scores for predicting outcome after microvascular decompression for trigeminal neuralgia. J Neurosurg 141:1056–1062. 10.3171/2024.1.JNS232053 [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Liu R, Deng Z, Zhang L et al (2020) The Long-Term Outcomes and Predictors of Microvascular Decompression with or without Partial Sensory Rhizotomy for Trigeminal Neuralgia. J Pain Res 13:301–312. 10.2147/JPR.S225188 [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Dong B, Xu R, Lim M (2023) The pathophysiology of trigeminal neuralgia: a molecular review. J Neurosurg 139:1471–1479. 10.3171/2023.2.JNS23274 [DOI] [PubMed] [Google Scholar]
17.Stern JI, Ali R, Chiang C-C, Robertson CE (2024) Pathophysiology and Management of Refractory Trigeminal Neuralgia. Curr Neurol Neurosci Rep 25:10. 10.1007/s11910-024-01387-2 [DOI] [PubMed] [Google Scholar]
18.Lin J, Zhang Y, Li W et al (2021) Flatness of the Meckel cave may cause primary trigeminal neuralgia: a radiomics-based study. J Headache Pain 22:104. 10.1186/s10194-021-01317-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Zhang F, Ni Y, Luo G et al (2024) Independent association of the Meckel’s cave with trigeminal neuralgia and development of a screening tool. Eur J Radiol 171:111272. 10.1016/j.ejrad.2023.111272 [DOI] [PubMed] [Google Scholar]
20.Li S, Liao C, Qian M et al (2022) Narrow ovale foramina may be involved in the development of primary trigeminal neuralgia. Front Neurol 13:1013216. 10.3389/fneur.2022.1013216 [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Zhong H, Zhang W, Sun S, Bie Y (2022) MRI Findings in Trigeminal Neuralgia without Neurovascular Compression: Implications of Petrous Ridge and Trigeminal Nerve Angles. Korean J Radiol 23:821–827. 10.3348/kjr.2021.0771 [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Di Stefano G, De Stefano G, Leone C et al (2020) Concomitant continuous pain in patients with trigeminal neuralgia is associated with trigeminal nerve root atrophy. Cephalalgia 40:1502–1510. 10.1177/0333102420949206 [DOI] [PubMed] [Google Scholar]
23.Li S, Liao C, Wang G et al (2023) Efficacy and Complications of Percutaneous Balloon Compression for Patients With Trigeminal Neuralgia With and Without Concomitant Continuous Pain. Pain Physician 26:E823–E832 [PubMed] [Google Scholar]
24.Wang J, Li M, Zhang Z et al (2025) Association between mental disorders and trigeminal neuralgia: a cohort study and Mendelian randomization analysis. J Headache Pain 26:74. 10.1186/s10194-025-02026-y [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Lee J-A, Han YK, Jung WJ et al (2025) Associations between personality traits and pain experiences in trigeminal neuralgia. J Headache Pain 26:68. 10.1186/s10194-025-02010-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
26.DosSantos MFH, Ramos LR, Bonacchi L et al (2026) Subcortical neuroimaging signature of painful disorders of the trigeminal nerve: a UK Biobank study. J Headache Pain 27:40. 10.1186/s10194-026-02271-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Chu T, Xiong F, Gao F et al (2026) Fixel–based analysis reveals distributed white matter alterations associated with pain severity in trigeminal neuralgia. J Headache Pain 27:84. 10.1186/s10194-026-02306-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Srisaikaew P, Latypov TH, Jörgens D et al (2026) Microstructural alterations of the human fornix in trigeminal neuralgia. J Pain 106325. 10.1016/j.jpain.2026.106325 [DOI] [PubMed]
29.Ide Y, Iwata K, Hayashi Y et al (2026) Trigeminal ganglion-trigeminal spinal subnucleus oralis pathway contributes to orofacial neuropathic pain. Pain 167:1331–1342. 10.1097/j.pain.0000000000003937 [DOI] [PubMed] [Google Scholar]
30.Yu D, Xie K (2024) A Predictive Model for the Risk of Postsurgery Pain Recurrence in the V1 Branch of the Trigeminal Nerve. Pain Physician 27:E147–E155 [PubMed] [Google Scholar]
31.Chen Z, Rui M, Xie K (2023) Nomogram Model for Predicting the Risk of Recurrence after Surgery of the V3 Branch of the Trigeminal Nerve. Pain Physician 26:E583–E590 [PubMed] [Google Scholar]
32.Olesen J (2018) International Classification of Headache Disorders. Lancet Neurol 17:396–397. 10.1016/S1474-4422(18)30085-1 [DOI] [PubMed] [Google Scholar]
33.Rogers CL, Shetter AG, Fiedler JA et al (2000) Gamma knife radiosurgery for trigeminal neuralgia: the initial experience of The Barrow Neurological Institute. Int J Radiat Oncol Biol Phys 47:1013–1019. 10.1016/s0360-3016(00)00513-7 [DOI] [PubMed] [Google Scholar]
34.Williams B, Mancia G, Spiering W et al (2018) 2018 ESC/ESH Guidelines for the management of arterial hypertension. Eur Heart J 39:3021–3104. 10.1093/eurheartj/ehy339 [DOI] [PubMed] [Google Scholar]
35.Maarbjerg S, Wolfram F, Gozalov A et al (2015) Significance of neurovascular contact in classical trigeminal neuralgia. Brain 138:311–319. 10.1093/brain/awu349 [DOI] [PubMed] [Google Scholar]
36.Wolff RF, Moons KGM, Riley RD et al (2019) PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies. Ann Intern Med 170:51–58. 10.7326/M18-1376 [DOI] [PubMed] [Google Scholar]
37.Collins GS, Moons KGM, Dhiman P et al (2024) TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385:e078378. 10.1136/bmj-2023-078378 [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Wu S, Chen J (2026) Is age-related myelinodegenerative change an initial risk factor of neurodegenerative diseases? Neural Regen Res 21:648–658. 10.4103/NRR.NRR-D-24-00848 [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Le Thuc O, García-Cáceres C (2024) Obesity-induced inflammation: connecting the periphery to the brain. Nat Metab 6:1237–1252. 10.1038/s42255-024-01079-8 [DOI] [PubMed] [Google Scholar]
40.Gouju J, Legeay S (2023) Pharmacokinetics of obese adults: Not only an increase in weight. Biomed Pharmacother 166:115281. 10.1016/j.biopha.2023.115281 [DOI] [PubMed] [Google Scholar]
41.Melzack R (1987) The short-form McGill Pain Questionnaire. Pain 30:191–197. 10.1016/0304-3959(87)91074-8 [DOI] [PubMed] [Google Scholar]
42.Li S, Liao C, Yang X, Zhang W (2023) Association of concomitant continuous pain in trigeminal neuralgia with a narrow foramen ovale. Front Neurol 14:1277654. 10.3389/fneur.2023.1277654 [DOI] [PMC free article] [PubMed] [Google Scholar]
43.Zhang Y, Mao Z, Pan L et al (2018) Dysregulation of Pain- and Emotion-Related Networks in Trigeminal Neuralgia. Front Hum Neurosci 12:107. 10.3389/fnhum.2018.00107 [DOI] [PMC free article] [PubMed] [Google Scholar]
44.Wang Y, Cao D, Remeniuk B et al (2017) Altered brain structure and function associated with sensory and affective components of classic trigeminal neuralgia. Pain 158:1561–1570. 10.1097/j.pain.0000000000000951 [DOI] [PubMed] [Google Scholar]
45.Tsai Y-H, Yuan R, Patel D et al (2018) Altered structure and functional connection in patients with classical trigeminal neuralgia. Hum Brain Mapp 39:609–621. 10.1002/hbm.23696 [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Zhang Y, Mao Z, Pan L et al (2019) Frequency-specific alterations in cortical rhythms and functional connectivity in trigeminal neuralgia. Brain Imaging Behav 13:1497–1509. 10.1007/s11682-019-00105-8 [DOI] [PubMed] [Google Scholar]
47.Zhang P, Jiang Y, Liu G et al (2021) Altered brain functional network dynamics in classic trigeminal neuralgia: a resting-state functional magnetic resonance imaging study. J Headache Pain 22:147. 10.1186/s10194-021-01354-z [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Yu X, Lv W, Liu X et al (2023) Abnormal activation of brain regions in idiopathic trigeminal neuralgia patients by fMRI: An activation likelihood estimation meta-analysis. Clin Neurol Neurosurg 228:107710. 10.1016/j.clineuro.2023.107710 [DOI] [PubMed] [Google Scholar]
49.Xu H, Seminowicz DA, Krimmel SR et al (2022) Altered Structural and Functional Connectivity of Salience Network in Patients with Classic Trigeminal Neuralgia. J Pain 23:1389–1399. 10.1016/j.jpain.2022.02.012 [DOI] [PubMed] [Google Scholar]
50.Zhang P, Wan X, Ai K et al (2022) Rich-club reorganization and related network disruptions are associated with the symptoms and severity in classic trigeminal neuralgia patients. Neuroimage Clin 36:103160. 10.1016/j.nicl.2022.103160 [DOI] [PMC free article] [PubMed] [Google Scholar]
51.Wang Z, Zhao Z, Song Z et al (2024) Functional alterations of the brain default mode network and somatosensory system in trigeminal neuralgia. Sci Rep 14:10205. 10.1038/s41598-024-60273-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Wang Z, Li Z, Zhou G et al (2024) Graph theory-driven structural and functional connectivity analyses revealing regulatory mechanisms of brain network in patients with classic trigeminal neuralgia. Brain Imaging Behav. 10.1007/s11682-024-00915-5 [DOI] [PubMed] [Google Scholar]
53.Neumann N, Domin M, Schmidt C-O, Lotze M (2023) Chronic pain is associated with less grey matter volume in the anterior cingulum, anterior and posterior insula and hippocampus across three different chronic pain conditions. Eur J Pain 27:1239–1248. 10.1002/ejp.2153 [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Material 1^{(35.8KB, docx)}

Supplementary Material 2^{(271KB, docx)}

Supplementary Material 3^{(31.6KB, docx)}

Data Availability Statement

No datasets were generated or analysed during the current study.

[CR1] 1.Ashina S, Robertson CE, Srikiatkhachorn A et al (2024) Trigeminal neuralgia. Nat Rev Dis Primers 10:39. 10.1038/s41572-024-00523-z [DOI] [PubMed] [Google Scholar]

[CR2] 2.Cruccu G, Di Stefano G, Truini A (2020) Trigeminal Neuralgia. N Engl J Med 383:754–762. 10.1056/NEJMra1914484 [DOI] [PubMed] [Google Scholar]

[CR3] 3.Bendtsen L, Zakrzewska JM, Heinskou TB et al (2020) Advances in diagnosis, classification, pathophysiology, and management of trigeminal neuralgia. Lancet Neurol 19:784–796. 10.1016/S1474-4422(20)30233-7 [DOI] [PubMed] [Google Scholar]

[CR4] 4.Tölle T, Dukes E, Sadosky A (2006) Patient burden of trigeminal neuralgia: results from a cross-sectional survey of health state impairment and treatment patterns in six European countries. Pain Pract 6:153–160. 10.1111/j.1533-2500.2006.00079.x [DOI] [PubMed] [Google Scholar]

[CR5] 5.Zakrzewska JM, Wu J, Mon-Williams M et al (2017) Evaluating the impact of trigeminal neuralgia. Pain 158:1166–1174. 10.1097/j.pain.0000000000000853 [DOI] [PubMed] [Google Scholar]

[CR6] 6.Spatz AL, Zakrzewska JM, Kay EJ (2007) Decision analysis of medical and surgical treatments for trigeminal neuralgia: how patient evaluations of benefits and risks affect the utility of treatment decisions. Pain 131:302–310. 10.1016/j.pain.2007.02.009 [DOI] [PubMed] [Google Scholar]

[CR7] 7.Bendtsen L, Zakrzewska JM, Abbott J et al (2019) European Academy of Neurology guideline on trigeminal neuralgia. Eur J Neurol 26:831–849. 10.1111/ene.13950 [DOI] [PubMed] [Google Scholar]

[CR8] 8.Rana MH, Khan AAG, Khalid I et al (2023) Therapeutic Approach for Trigeminal Neuralgia: A Systematic Review. Biomedicines 11:2606. 10.3390/biomedicines11102606 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR9] 9.Pergolizzi JV, LeQuang JA, El-Tallawy SN et al (2024) An update on pharmacotherapy for trigeminal neuralgia. Expert Rev Neurother 24:773–786. 10.1080/14737175.2024.2365946 [DOI] [PubMed] [Google Scholar]

[CR10] 10.Gronseth G, Cruccu G, Alksne J et al (2008) Practice parameter: the diagnostic evaluation and treatment of trigeminal neuralgia (an evidence-based review): report of the Quality Standards Subcommittee of the American Academy of Neurology and the European Federation of Neurological Societies. Neurology 71:1183–1190. 10.1212/01.wnl.0000326598.83183.04 [DOI] [PubMed] [Google Scholar]

[CR11] 11.Li S, Chen H, Chen J et al (2024) Predicting long-term outcomes in patients with classical trigeminal neuralgia following microvascular decompression with an MRI-based radiomics nomogram: a multicentre study. Eur Radiol. 10.1007/s00330-024-10775-8 [DOI] [PubMed] [Google Scholar]

[CR12] 12.Andersen ASS, Heinskou TB, Rochat P et al (2022) Microvascular decompression in trigeminal neuralgia - a prospective study of 115 patients. J Headache Pain 23:145. 10.1186/s10194-022-01520-x [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR13] 13.Heinskou TB, Rochat P, Maarbjerg S et al (2019) Prognostic factors for outcome of microvascular decompression in trigeminal neuralgia: A prospective systematic study using independent assessors. Cephalalgia 39:197–208. 10.1177/0333102418783294 [DOI] [PubMed] [Google Scholar]

[CR14] 14.Xie ME, Halbert-Elliott K, Nair SK et al (2024) External assessment of preoperative scores for predicting outcome after microvascular decompression for trigeminal neuralgia. J Neurosurg 141:1056–1062. 10.3171/2024.1.JNS232053 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR15] 15.Liu R, Deng Z, Zhang L et al (2020) The Long-Term Outcomes and Predictors of Microvascular Decompression with or without Partial Sensory Rhizotomy for Trigeminal Neuralgia. J Pain Res 13:301–312. 10.2147/JPR.S225188 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR16] 16.Dong B, Xu R, Lim M (2023) The pathophysiology of trigeminal neuralgia: a molecular review. J Neurosurg 139:1471–1479. 10.3171/2023.2.JNS23274 [DOI] [PubMed] [Google Scholar]

[CR17] 17.Stern JI, Ali R, Chiang C-C, Robertson CE (2024) Pathophysiology and Management of Refractory Trigeminal Neuralgia. Curr Neurol Neurosci Rep 25:10. 10.1007/s11910-024-01387-2 [DOI] [PubMed] [Google Scholar]

[CR18] 18.Lin J, Zhang Y, Li W et al (2021) Flatness of the Meckel cave may cause primary trigeminal neuralgia: a radiomics-based study. J Headache Pain 22:104. 10.1186/s10194-021-01317-4 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR19] 19.Zhang F, Ni Y, Luo G et al (2024) Independent association of the Meckel’s cave with trigeminal neuralgia and development of a screening tool. Eur J Radiol 171:111272. 10.1016/j.ejrad.2023.111272 [DOI] [PubMed] [Google Scholar]

[CR20] 20.Li S, Liao C, Qian M et al (2022) Narrow ovale foramina may be involved in the development of primary trigeminal neuralgia. Front Neurol 13:1013216. 10.3389/fneur.2022.1013216 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR21] 21.Zhong H, Zhang W, Sun S, Bie Y (2022) MRI Findings in Trigeminal Neuralgia without Neurovascular Compression: Implications of Petrous Ridge and Trigeminal Nerve Angles. Korean J Radiol 23:821–827. 10.3348/kjr.2021.0771 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR22] 22.Di Stefano G, De Stefano G, Leone C et al (2020) Concomitant continuous pain in patients with trigeminal neuralgia is associated with trigeminal nerve root atrophy. Cephalalgia 40:1502–1510. 10.1177/0333102420949206 [DOI] [PubMed] [Google Scholar]

[CR23] 23.Li S, Liao C, Wang G et al (2023) Efficacy and Complications of Percutaneous Balloon Compression for Patients With Trigeminal Neuralgia With and Without Concomitant Continuous Pain. Pain Physician 26:E823–E832 [PubMed] [Google Scholar]

[CR24] 24.Wang J, Li M, Zhang Z et al (2025) Association between mental disorders and trigeminal neuralgia: a cohort study and Mendelian randomization analysis. J Headache Pain 26:74. 10.1186/s10194-025-02026-y [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR25] 25.Lee J-A, Han YK, Jung WJ et al (2025) Associations between personality traits and pain experiences in trigeminal neuralgia. J Headache Pain 26:68. 10.1186/s10194-025-02010-6 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR26] 26.DosSantos MFH, Ramos LR, Bonacchi L et al (2026) Subcortical neuroimaging signature of painful disorders of the trigeminal nerve: a UK Biobank study. J Headache Pain 27:40. 10.1186/s10194-026-02271-9 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR27] 27.Chu T, Xiong F, Gao F et al (2026) Fixel–based analysis reveals distributed white matter alterations associated with pain severity in trigeminal neuralgia. J Headache Pain 27:84. 10.1186/s10194-026-02306-1 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR28] 28.Srisaikaew P, Latypov TH, Jörgens D et al (2026) Microstructural alterations of the human fornix in trigeminal neuralgia. J Pain 106325. 10.1016/j.jpain.2026.106325 [DOI] [PubMed]

[CR29] 29.Ide Y, Iwata K, Hayashi Y et al (2026) Trigeminal ganglion-trigeminal spinal subnucleus oralis pathway contributes to orofacial neuropathic pain. Pain 167:1331–1342. 10.1097/j.pain.0000000000003937 [DOI] [PubMed] [Google Scholar]

[CR30] 30.Yu D, Xie K (2024) A Predictive Model for the Risk of Postsurgery Pain Recurrence in the V1 Branch of the Trigeminal Nerve. Pain Physician 27:E147–E155 [PubMed] [Google Scholar]

[CR31] 31.Chen Z, Rui M, Xie K (2023) Nomogram Model for Predicting the Risk of Recurrence after Surgery of the V3 Branch of the Trigeminal Nerve. Pain Physician 26:E583–E590 [PubMed] [Google Scholar]

[CR32] 32.Olesen J (2018) International Classification of Headache Disorders. Lancet Neurol 17:396–397. 10.1016/S1474-4422(18)30085-1 [DOI] [PubMed] [Google Scholar]

[CR33] 33.Rogers CL, Shetter AG, Fiedler JA et al (2000) Gamma knife radiosurgery for trigeminal neuralgia: the initial experience of The Barrow Neurological Institute. Int J Radiat Oncol Biol Phys 47:1013–1019. 10.1016/s0360-3016(00)00513-7 [DOI] [PubMed] [Google Scholar]

[CR34] 34.Williams B, Mancia G, Spiering W et al (2018) 2018 ESC/ESH Guidelines for the management of arterial hypertension. Eur Heart J 39:3021–3104. 10.1093/eurheartj/ehy339 [DOI] [PubMed] [Google Scholar]

[CR35] 35.Maarbjerg S, Wolfram F, Gozalov A et al (2015) Significance of neurovascular contact in classical trigeminal neuralgia. Brain 138:311–319. 10.1093/brain/awu349 [DOI] [PubMed] [Google Scholar]

[CR36] 36.Wolff RF, Moons KGM, Riley RD et al (2019) PROBAST: A Tool to Assess the Risk of Bias and Applicability of Prediction Model Studies. Ann Intern Med 170:51–58. 10.7326/M18-1376 [DOI] [PubMed] [Google Scholar]

[CR37] 37.Collins GS, Moons KGM, Dhiman P et al (2024) TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385:e078378. 10.1136/bmj-2023-078378 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR38] 38.Wu S, Chen J (2026) Is age-related myelinodegenerative change an initial risk factor of neurodegenerative diseases? Neural Regen Res 21:648–658. 10.4103/NRR.NRR-D-24-00848 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR39] 39.Le Thuc O, García-Cáceres C (2024) Obesity-induced inflammation: connecting the periphery to the brain. Nat Metab 6:1237–1252. 10.1038/s42255-024-01079-8 [DOI] [PubMed] [Google Scholar]

[CR40] 40.Gouju J, Legeay S (2023) Pharmacokinetics of obese adults: Not only an increase in weight. Biomed Pharmacother 166:115281. 10.1016/j.biopha.2023.115281 [DOI] [PubMed] [Google Scholar]

[CR41] 41.Melzack R (1987) The short-form McGill Pain Questionnaire. Pain 30:191–197. 10.1016/0304-3959(87)91074-8 [DOI] [PubMed] [Google Scholar]

[CR42] 42.Li S, Liao C, Yang X, Zhang W (2023) Association of concomitant continuous pain in trigeminal neuralgia with a narrow foramen ovale. Front Neurol 14:1277654. 10.3389/fneur.2023.1277654 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR43] 43.Zhang Y, Mao Z, Pan L et al (2018) Dysregulation of Pain- and Emotion-Related Networks in Trigeminal Neuralgia. Front Hum Neurosci 12:107. 10.3389/fnhum.2018.00107 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR44] 44.Wang Y, Cao D, Remeniuk B et al (2017) Altered brain structure and function associated with sensory and affective components of classic trigeminal neuralgia. Pain 158:1561–1570. 10.1097/j.pain.0000000000000951 [DOI] [PubMed] [Google Scholar]

[CR45] 45.Tsai Y-H, Yuan R, Patel D et al (2018) Altered structure and functional connection in patients with classical trigeminal neuralgia. Hum Brain Mapp 39:609–621. 10.1002/hbm.23696 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR46] 46.Zhang Y, Mao Z, Pan L et al (2019) Frequency-specific alterations in cortical rhythms and functional connectivity in trigeminal neuralgia. Brain Imaging Behav 13:1497–1509. 10.1007/s11682-019-00105-8 [DOI] [PubMed] [Google Scholar]

[CR47] 47.Zhang P, Jiang Y, Liu G et al (2021) Altered brain functional network dynamics in classic trigeminal neuralgia: a resting-state functional magnetic resonance imaging study. J Headache Pain 22:147. 10.1186/s10194-021-01354-z [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR48] 48.Yu X, Lv W, Liu X et al (2023) Abnormal activation of brain regions in idiopathic trigeminal neuralgia patients by fMRI: An activation likelihood estimation meta-analysis. Clin Neurol Neurosurg 228:107710. 10.1016/j.clineuro.2023.107710 [DOI] [PubMed] [Google Scholar]

[CR49] 49.Xu H, Seminowicz DA, Krimmel SR et al (2022) Altered Structural and Functional Connectivity of Salience Network in Patients with Classic Trigeminal Neuralgia. J Pain 23:1389–1399. 10.1016/j.jpain.2022.02.012 [DOI] [PubMed] [Google Scholar]

[CR50] 50.Zhang P, Wan X, Ai K et al (2022) Rich-club reorganization and related network disruptions are associated with the symptoms and severity in classic trigeminal neuralgia patients. Neuroimage Clin 36:103160. 10.1016/j.nicl.2022.103160 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR51] 51.Wang Z, Zhao Z, Song Z et al (2024) Functional alterations of the brain default mode network and somatosensory system in trigeminal neuralgia. Sci Rep 14:10205. 10.1038/s41598-024-60273-6 [DOI] [PMC free article] [PubMed] [Google Scholar]

[CR52] 52.Wang Z, Li Z, Zhou G et al (2024) Graph theory-driven structural and functional connectivity analyses revealing regulatory mechanisms of brain network in patients with classic trigeminal neuralgia. Brain Imaging Behav. 10.1007/s11682-024-00915-5 [DOI] [PubMed] [Google Scholar]

[CR53] 53.Neumann N, Domin M, Schmidt C-O, Lotze M (2023) Chronic pain is associated with less grey matter volume in the anterior cingulum, anterior and posterior insula and hippocampus across three different chronic pain conditions. Eur J Pain 27:1239–1248. 10.1002/ejp.2153 [DOI] [PubMed] [Google Scholar]

PERMALINK

Exploratory development of prediction models for pharmacotherapy outcomes in trigeminal neuralgia: a combined analysis based on multi-source data

Hongan Yang

Baijintao Sun

Mingtao Li

Xuezhao Yang

Mingyan Li

Chuan Zhang

Hanfeng Yang

Abstract

Background

Methods

Results

Conclusion

Supplementary Information

Introduction

Methods

Study population screening and study design

Fig. 1.

DRTN grouping criteria and detail

Data collection

Table 1.

Table 2.

Table 3.

Feature selection

Fig. 2.

Fig. 4.

Machine learning model construction and evaluation

Fig. 3.

Cox regression model construction and evaluation

Statistical analysis

Results

Intergroup feature comparison and feature selection in the primary cohort

Machine learning model performance

Table 4.

SHAP analysis of the optimal model (SVM-RBF)

Cox regression models

Univariate and multivariate regression analysis

Table 5.

Table 6.

Cox model performance and time-dependent discrimination analysis

Discussion

Multi-domain feature interpretation

Machine learning model performance analysis

Cox model performance analysis

Limitation

Fig. 5.

Conclusion

Supplementary Information

Abbreviations

Author contributions

Funding

Data availability

Declarations

Competing interests

Footnotes

Contributor Information

References

Associated Data

Supplementary Materials

Data Availability Statement

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases