Haematologic Data Improves Long‐Term Prediction Accuracy of Artificial Intelligence Models for Temporomandibular Disorders

ABSTRACT

Objectives

This study aimed to develop and evaluate an artificial intelligence (AI) model to predict long‐term treatment outcomes in temporomandibular disorder (TMD) patients using clinical data and verify the value of adding haematologic data in enhancing predictive accuracy.

Methods

The medical records of 132 TMD patients who visited the clinic and underwent 6 months of non‐invasive conservative treatment between 2013 and 2019 were included in this study. The clinical data and haematologic features were collected from medical records. A decision tree algorithm was employed for feature selection, followed by a deep neural network (DNN) to build the prediction model. The performance of the models based on the decision tree algorithm and DNN was evaluated.

Results

The decision tree model achieved an accuracy of 90.6% and an F1‐score of 0.800. The subjective pain‐related features, along with haematologic markers associated with systemic inflammation, were proven to be important features in the decision tree model. The predictive performance of the DNN model improved as haematologic features were added, with the final model achieving an accuracy of 90.6% and an F1‐score of 0.769.

Conclusions

This study showed the potential of machine learning models in predicting long‐term TMD prognosis using clinical and haematological features. In addition, these findings highlight the importance of including both subjective pain assessments and systemic haematologic markers for the development of aetiology‐based diagnostic systems for TMD to enhance clinical decision‐making and prognosis prediction accuracy.

1. Introduction

Temporomandibular disorders (TMD) refers to a complex condition characterised by pain and dysfunction of the masticatory system including the temporomandibular joints (TMJ) and masticatory muscles [1]. TMD affects a substantial portion of the general population, with studies indicating a prevalence rate of 6% to 12%. It is twice as commonly observed in females and particularly prevalent among individuals aged 20 to 40 years [2]. While the primary symptoms include jaw pain, joint noises and restricted jaw movement, more complex forms may be associated with headache, other bodily pain and sleep problems. TMD significantly impacts quality of life, particularly when severe pain and functional limitations are present, leading to substantial discomfort in daily activities [3].

TMD treatment primarily consists of conservative approaches, including physical therapy, medication, cognitive behavioural therapy and stabilisation splints [4]. The majority of patients achieve pain relief and functional recovery through such treatment. However, some patients do not respond well to conventional treatment, resulting in chronic pain and prolonged treatment [5]. This not only increases the disease burden but also significantly deteriorates the patient's quality of life [6]. Additionally, repeated treatment failures may negatively impact overall prognosis [7]. Therefore, it would be ideal to predict the possibility of chronic TMD in advance to optimise treatment plans that will result in enhanced outcomes.

The aetiology of TMD is multifactorial, involving anatomical, physiological, psychological and environmental factors [8]. Various studies have been conducted to locate strong predictors of chronic TMD. It has been suggested that psychological factors such as somatization and depression may play significant roles in recalcitrant TMD [9, 10]. Additionally, widespread pain has been associated with worse treatment response [11, 12], indicating a possible role of central sensitization. Recently, haematological markers of systemic inflammation have been implicated in the pain chronicity of conditions such as fibromyalgia and complex regional pain syndrome [13, 14], and similar immune dysregulation has been suggested in TMD patients [15, 16]. However, previous studies face limitations in predicting pain chronicity in TMD patients due to their reliance on statistical differences based on a limited number of factors. This highlights the need for a more comprehensive and simultaneous analysis of various factors.

Artificial intelligence (AI) excels at analysing large datasets and identifying complex patterns. In the medical field, AI has shown significant potential in diagnosis and prognosis prediction. Therefore, the purpose of this study was to develop an AI model that predicts the long‐term treatment outcomes of TMD patients using clinical data and evaluate the value of additional haematologic data in increasing diagnostic accuracy. By providing a tool to identify high‐risk patients of chronic TMD at an early stage, we aim to manage TMD patients based on personalised treatment approaches and ultimately assist in improving their prognosis and quality of life.

2. Materials and Methods

2.1. Participants and Data Collection

This study is based on medical records of patients who visited the Department of Oral Medicine, Seoul National University Dental Hospital with symptoms of TMD between March 2013 and April 2019. This patient data was part of a dataset included in a previous study [17]. This study was approved by the Institutional Review Board (IRB) of Seoul National University Dental Hospital (ERI24024). All patients provided informed consent on the academic usage of their clinical data on the first visit to the hospital and the requirement for additional informed consent was waived based on the retrospective nature of the study. All study procedures adhered to the Declaration of Helsinki and relevant guidelines.

During their initial visit, all participants underwent structured interviews, medical and dental history taking, intraoral examinations, TMD assessment based on the research diagnostic criteria for TMD (RDC/TMD) [18], psychological evaluation with RDC/TMD axis II questionnaires and Symptom Checklist‐90‐Revised (SCL‐90‐R), radiographic examinations and haematologic tests. The laboratory tests included complete blood count (CBC) with white blood cell (WBC) differential, red blood cell (RBC) indices, and blood chemistry along with C‐reactive protein (CRP), antinuclear antibody (ANA) and rheumatoid factor (RF). Haematologic biomarkers of systemic inflammation (NLR. neutrophil/lymphocyte count), derived NLR (dNLR, absolute neutrophil/[white blood cell‐absolute neutrophil count]), LMR (lymphocyte/monocyte count), PLR (platelet/lymphocyte count) and systemic immune‐inflammation index (SII, platelet × [neutrophil/lymphocyte count]) were calculated [19, 20]. All diagnoses were made by a board‐certified oral medicine specialist (J.W.P.) with over 15 years of clinical experience in orofacial pain diagnosis and treatment.

All participants received conservative treatment including control of contributing factors, self‐exercise, occlusal stabilisation splints, physical therapy and medications such as non‐steroidal anti‐inflammatory drugs. At each visit, unassisted opening without pain, maximum unassisted opening, presence of pain on the palpation of masticatory muscles and TMJ capsule areas and subjective pain intensity were recorded. Subjective pain intensity was measured using the numeric rating scale (NRS, 0–10, with 10 representing worst pain). A total of 132 patients who completed the 6‐month treatment period without missing clinical data were included in the final analysis and categorised as improved or unimproved groups based on their long‐term treatment response. The improved group was defined as those with a reduction in NRS score of 2 or more points after 6 months of treatment compared to baseline; the unimproved group included those with less than a 2‐point reduction. Ultimately, 95 and 37 patients were assigned to improved and unimproved groups, respectively.

2.2. Feature Selection and Data Preprocessing

The clinical and haematologic data collected during the initial visit before treatment initiation were used for analysis. Clinical data included patient age and sex, duration and intensity of pain, TMD axis I classification, psychological characteristics, comorbid symptoms in other body regions (headache, neck, shoulder, low back, leg, and arm pain and gastrointestinal symptoms), sleep disturbance, unassisted opening without pain, maximum unassisted opening, capsule palpation, masticatory muscle and neck muscle palpation response, presence of tooth attrition, tongue and mucosal ridging, pain on opening and eccentric movement and degenerative changes of TMJ condyles. Haematological data included CBC with WBC differential, erythrocyte sedimentation rate (ESR), ANA, RA and inflammatory biomarkers. In total, 89 features were used by combining clinical and haematologic data, and the treatment response was labelled as 0 or 1 based on improved or unimproved group.

Data preprocessing which is essential for the quality and reliability of analysis included data cleaning to remove errors, and data transforming to map data for easier interpretation. The outliers were transformed using interquartile range (IQR). For each feature value, $Q_1$ and $Q_3$ were set as the first and third quartile and defined $\mathrm{IQR}=Q_3-Q_1$. Setting an upper bound $b_u=Q_3+1.5\times \mathrm{IQR}$ and a lower bound $b_l=Q_1-1.5\times \mathrm{IQR}$, the values exceeding the bounds were replaced with $b_u$ and $b_l$, accordingly.

2.3. Construction of Prediction Model by AI

The model development process consisted of two stages. In the first stage, a decision tree algorithm was employed to extract significant features for label prediction [21]. This step aimed to identify the most relevant clinical and haematologic features contributing to treatment response. The decision tree was trained on 100 data samples and tested on 32 data samples, with a maximum depth of 20 and a minimum of 8 samples required to split a node.

In the second stage, a deep neural network (DNN) was used to construct the actual prediction model utilising the features extracted in the first stage [22]. The DNN architecture was designed as y = f(Wx), where the input x consisted of the top important variables identified by the decision tree, and the output y ∈ R ¹. The learnable parameter was denoted by W, and the non‐linear function f was represented by the widely‐used ReLU activation function. The dataset was split into two subsets of training (100 samples) and testing (32 samples) while maintaining the labels' ratio. A standard scaler was applied using the mean (μ) and standard deviation (SD, σ) of the training data. The model utilised a binary cross‐entropy (BCE) loss function and was optimised using the Adam optimiser. The hyperparameters for the DNN were set as follows: Epochs: 25; Learning Rate: 0.01; Batch Size: 15; Seed: 41.

3. Results

3.1. Clinical Characteristics and Haematologic Markers According to Prognosis Group

The clinical characteristics of the improved (n = 95) and unimproved (n = 37) groups are summarised in Table 1. The mean age and gender distribution of patients in the improved and unimproved groups showed no significant differences. In terms of pain characteristics, pain intensity measured by NRS was significantly higher in the improved group (5.12 ± 1.76) than in the unimproved group (3.57 ± 1.99) (p < 0.001). Additionally, the improved group exhibited significantly more pain on mouth opening (p = 0.005) and during eccentric movement (p = 0.001) compared to the unimproved group before treatment. However, the range of mouth opening (unassisted opening without pain and maximum unassisted opening), pain on capsule and muscle palpation, and duration and origin of pain did not differ significantly between the groups. Other clinical characteristics, including diagnosis type, disability level, comorbid symptoms, radiographic findings and psychological characteristics, did not show statistically significant differences between the groups.

TABLE 1.

Clinical characteristics according to long‐term prognosis group before treatment, mean ± SD, n (%).

	Improved (n = 95)	Unimproved (n = 37)	p
Clinical
Age (years)	30.59 ± 10.72	30.22 ± 11.56	0.614
Gender
Male	14 (14.7)	2 (5.4)	0.140
Female	81 (85.3)	35 (94.6)
Pain duration (months)	27.29 ± 43.28	23.67 ± 27.22	0.433
RDC/TMD axis I diagnosis
Myofascial pain	63 (66.3)	21 (56.8)	0.305
Disc displacement	84 (88.4)	29 (78.4)	0.140
Arthalgia/osteoarthritis/osteoarthrosis	76 (80.0)	26 (70.3)	0.231
Characteristic pain intensity	46.22 ± 18.19	39.27 ± 19.63	0.064
Disability days	34.87 ± 54.62	48.11 ± 62.47	0.194
GCPS
Low disability	51 (53.7)	17 (45.9)	0.424
High disability	44 (46.4)	20 (54.1)
Pain intensity (NRS)	5.12 ± 1.76	3.57 ± 1.99	> 0.001*
Unassisted opening without pain (mm)	36.48 ± 10.43	39.59 ± 9.98	0.118
Maximum unassisted opening (mm)	42.44 ± 8.37	44.27 ± 7.58	0.344
Pain on mouth opening	64 (67.4)	15 (40.5)	0.005*
Pain on eccentric movement	1.01 ± 1.03	0.41 ± 0.76	0.001*
Pain on capsule palpation	47 (49.5)	14 (37.8)	0.228
Pain on Masticatory muscle palpation	66 (69.5)	24 (64.9)	0.610
Pain on cervical muscle palpation	44 (46.3)	16 (43.2)	0.750
Tongue ridging	76 (80.0)	30 (81.1)	0.888
Mucosal ridging	87 (91.6)	37 (100.0)	0.105
Comorbidities
Headache	46 (48.4)	20 (54.1)	0.561
Sleep disturbance	29 (30.5)	12 (32.4)	0.832
Neck and shoulder pain	56 (58.9)	22 (59.5)	0.957
Low back pain	40 (42.1)	12 (32.4)	0.307
Leg and arm pain	15 (15.8)	4 (10.8)	0.464
Gastrointestinal disorders	23 (24.2)	10 (27.0)	0.737
DJD on radiograph	58 (61.1)	22 (59.5)
RDC‐DEP	0.68 ± 0.66	0.75 ± 0.70	0.549
RDC‐SOM	0.78 ± 0.70	0.69 ± 0.56	0.800
RDC‐PSOM	0.62 ± 0.65	0.57 ± 0.57	0.941
SCL‐90‐R
Somatization	45.56 ± 7.91	47.08 ± 8.62	0.346
Obsessive‐compulsive	43.54 ± 9.19	42.51 ± 7.97	0.757
Interpersonal sensitivity	43.92 ± 10.02	42.57 ± 7.74	0.749
Depression	43.13 ± 8.44	42.49 ± 8.63	0.621
Anxiety	43.53 ± 7.56	44.41 ± 8.35	0.851
Hostility	44.12 ± 7.18	44.19 ± 8.21	0.939
Phobic anxiety	44.92 ± 7.50	46.19 ± 10.88	0.879
Paranoid ideation	42.38 ± 6.79	41.19 ± 5.29	0.584
Psychoticism	42.89 ± 6.13	42.51 ± 6.42	0.781

Abbreviations: DJD, degenerative joint disease of temporomandibular joint; GCPS, graded chronic pain scale; NRS, numeric rating scale; RDC‐DEP, depression score of RDC/TMD axis II; RDC‐PSOM, somatization score of RDC/TMD axis II without pain items; RDC‐SOM, somatization score of RDC/TMD axis II; RDC/TMD, research diagnostic criteria for temporomandibular disorders; SCL‐90‐R, symptom checklist‐90‐revised.

^* p < 0.05.

The haematologic markers of both groups are also shown in Table 2. The improved group had a significantly higher haemoglobin level (13.7 ± 1.1 g/dL) compared to the unimproved group (13.3 ± 1.2 g/dL) (p = 0.031). Additionally, the improved group showed a higher haematocrit and lower ESR compared to the unimproved group, but these two markers only demonstrated borderline significance (haematocrit, 40.5% ± 3.1% vs. 39.5% ± 3.2%, p = 0.073; ESR, 8.74 ± 7.60 mm/h vs. 10.46 ± 6.46 mm/h, p = 0.056). In the results of inflammatory biomarkers, more patients in the unimproved group exhibited abnormally low LMR values and abnormally high PLR values. The improved group had higher NLR, dNLR and PLR values and lower LMR values compared to the unimproved group; however, such differences did not reach statistical significance.

TABLE 2.

Haematological characteristics according to long‐term prognosis group before treatment, mean ± SD, n (%).

	Improved (n = 95)	Unimproved (n = 37)	p
WBC (103/μL)	6.08 ± 1.80	6.27 ± 1.38	0.331
RBC (106/μL)	4.45 ± 0.41	4.41 ± 0.43	0.103
Hgb (g/dL)	13.7 ± 1.1	13.3 ± 1.2	0.031*
Hct (%)	40.5 ± 3.1	39.5 ± 3.2	0.073
MCV (fL)	89.6 ± 3.7	89.7 ± 3.6	0.650
MCH (pg)	30.3 ± 1.3	30.1 ± 1.4	0.652
MCHC (g/dL)	33.8 ± 0.9	33.6 ± 0.9	0.085
Platelet (103/μL)	260.5 ± 71.4	260.2 ± 49.4	0.458
Total protein (g/dL)	7.58 ± 0.34	7.67 ± 0.37	0.190
ESR (mm/h)	8.74 ± 7.60	10.46 ± 6.46	0.056
ESR group (≥ 20.0 mm/h)	8 (8.4)	3 (8.1)	1.000
CRP (mg/dL)	0.09 ± 0.14	0.11 ± 0.22	0.884
CRP group (≥ 50 mg/dL)	3 (3.2)	2 (5.4)	0.619
NLR	1.97 ± 0.98	1.86 ± 0.82	0.588
NLR group (≥ 1.662 (F), ≥ 1.634 (M))	57 (60.0)	19 (51.4)	0.367
dNLR	1.47 ± 0.65	1.43 ± 0.64	0.753
LMR	4.51 ± 1.43	4.91 ± 1.46	0.119
LMR group (≤ 5.598 (F), ≤ 5.048 (M))	78 (82.1)	24 (64.9)	0.034*
PLR	144.5 ± 55.2	132.9 ± 38.9	0.459
PLR group (≥ 142.76 (F), ≥ 122.73 (M))	46 (48.4)	10 (27.0)	0.025*
RF positivity	7 (7.4)	1 (2.7)	0.441
FANA positivity	12 (12.6)	5 (13.5)	0.892

Abbreviations: CRP, C‐reactive protein; dNLR, derived NLR ratio; ESR, erythrocyte sedimentation rate; FANA, fluorescent antinuclear antibody; Hct, haematocrit; Hgb, haemoglobin; LMR, lymphocyte‐to‐monocyte ratio; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; MCV, mean corpuscular volume; MPV, mean platelet volume; NLR, neutrophil‐to‐lymphocyte ratio; PCT, plateletcrit; PLR, platelet‐to‐lymphocyte ratio; RBC, red blood cell; RF, rheumatoid factor; WBC, white blood cell.

^* p < 0.05.

3.2. Changes in Clinical Signs After 6 Months of Treatments According to Prognosis Group

In the improved group, all clinical signs significantly improved after 6 months of conservative treatment. Unassisted opening without pain and maximum unassisted opening increased, pain intensity decreased, and the proportion of patients experiencing pain on mouth opening, capsule and muscle palpation also significantly decreased. In contrast, the patients in the unimproved group experienced no improvement or a worsening of clinical signs, with no significant differences observed before and after treatment (Table 3).

TABLE 3.

Changes in clinical signs 6 months after treatment according to long‐term prognosis group, mean ± SD, n (%).

	Baseline	6 months	p
Improved (n = 95)
Pain intensity (NRS)	5.12 ± 1.76	1.23 ± 1.28	> 0.001*
Unassisted opening without pain (mm)	36.48 ± 10.43	43.09 ± 8.47	> 0.001*
Maximum unassisted opening (mm)	42.44 ± 8.37	44.80 ± 7.06	> 0.001*
Pain on mouth opening	64 (67.4)	32 (33.7)	> 0.001*
Pain on capsule palpation	47 (49.5)	32 (33.7)	0.027*
Pain on masticatory muscle palpation	66 (69.5)	37 (38.9)	> 0.001*
Unimproved (n = 37)
Pain intensity (NRS)	3.57 ± 1.99	3.76 ± 2.13	0.721
Unassisted opening without pain (mm)	39.59 ± 9.98	40.43 ± 8.55	0.557
Maximum unassisted opening (mm)	44.27 ± 7.58	43.16 ± 6.48	0.667
Pain on mouth opening	15 (40.5)	19 (59.5)	0.351
Pain on capsule palpation	14 (37.8)	13 (35.1)	0.809
Pain on masticatory muscle palpation	24 (64.9)	15 (40.5)	0.062

Abbreviation: NRS, numeric rating scale.

^* p < 0.05.

3.3. Prediction Model Performance and Feature Selection Using the Decision Tree Algorithm

The decision tree algorithm was applied to identify key features from a set of 89 available clinical and haematologic variables. The model generated from this process demonstrated a high level of performance, with an accuracy of 0.906, precision of 0.750, recall of 0.857 and an F1‐score of 0.800, as shown in the first row of Table 4.

TABLE 4.

Long‐term prognosis prediction performance of AI models.

	Accuracy	Precision	Recall	F1‐score
Decision tree algorithm	0.906	0.750	0.857	0.800
Deep neural network
Model 1 a	0.875	1.000	0.200	0.333
Model 2 b	0.781	0.462	1.000	0.632
Model 3 c	0.906	0.833	0.714	0.769

Abbreviations: ESR, erythrocyte sedimentation rate; Hgb, haemoglobin; LMR, lymphocyte‐to‐monocyte ratio; MCHC, mean corpuscular haemoglobin concentration; NRS, numeric rating scale.

^a Deep neural network model utilising two features (Pain intensity (NRS) and pain on eccentric movement).

^b Deep neural network model utilising five features (Pain intensity (NRS), pain on eccentric movement, LMR, MCHC and Hgb).

^c Deep neural network model utilising six features (Pain intensity (NRS), pain on eccentric movement, LMR, MCHC, Hgb and ESR).

Feature importance was calculated from the decision tree algorithm model. Pre‐treatment pain intensity (NRS) was the most important predictor in the prediction model (feature importance, FI, 0.129684). Moreover, two other pain‐related features—pain on eccentric movement (FI, 0.052795) and characteristic pain intensity score (FI, 0.039981)—were the second and fourth most important predictors, respectively. However, other clinical characteristics, including psychological characteristics, were not included in the top 10 most informative features. The remaining seven most informative features were haematologic markers (mean corpuscular haemoglobin concentration [MCHC, FI, 0.047409], Haemoglobin [Hgb, FI, 0.037668], Haematocrit [Hct, FI, 0.031650], LMR [FI, 0.029459], lymphocyte [FI, 0.025581], ESR [FI, 0.024667] and monocyte levels [FI, 0.024266]) (Figure 1).

FIGURE 1.

Importance of the top 10 features for long‐term prognosis prediction identified by the decision tree algorithm. ESR, erythrocyte sedimentation rate; Hct, haematocrit; Hgb, haemoglobin; LMR, lymphocyte‐to‐monocyte ratio; MCHC, mean corpuscular haemoglobin concentration; NRS, numeric rating scale.

3.4. Prediction Model Performance Using DNN

Among the top 10 features identified by the decision tree algorithm, highly correlated features were excluded and six were selected to develop the DNN‐based prediction model. The final feature set included pre‐treatment pain intensity, pain on eccentric movement, MCHC, Hgb concentration, LMR and ESR. These features were sequentially analysed in three stages to assess the impact of adding additional features on the model's performance.

Model 1 was developed using only the two pain‐related features (pain intensity and pain on eccentric movement). This model focused on evaluating how well the DNN could predict long‐term prognosis based on simple pain characteristics alone. Model 2 expanded upon Model 1 by incorporating three additional haematologic markers (MCHC, Hgb and LMR), resulting in a total of five features. Finally, Model 3 included all six selected features by adding ESR to the previous five features. This allowed us to assess the contribution of these haematologic and inflammatory markers to the model's predictive capability.

The performance of these models is summarised in Table 4. Model 1, which used only the two pain‐related features, achieved a relatively low F1‐score of 0.333. Model 2, with the inclusion of three haematologic markers, demonstrated improved performance, with an F1‐score of 0.632. Model 3, which utilised all six features, achieved the highest F1‐score of 0.769.

The Figure 2 displays the receiver operating characteristic (ROC)‐area under the curves (AUC) for the DNN‐based models using three different feature sets. It demonstrates that as more haematologic features are incorporated, the predictive performance of the models generally improves, especially with the final model showing the best discriminatory ability.

FIGURE 2.

ROC‐AUC for the DNN‐based (a) Model 1, (b) Model 2 and (c) Model 3. DNN, deep neural network; ROC‐AUC, receiver operating characteristic—area under the curve.

For each case in the test dataset, the predicted probabilities generated by the DNN‐based models were calculated, indicating the likelihood of classification into the unimproved group. Green bars represent cases from the improved group, while red bars represent cases from the unimproved group. It shows that the addition of more features resulted in improved prediction probabilities (Figure 3).

FIGURE 3.

Predicted probabilities for each case in the test dataset were generated by the DNN‐based (a) Model 1, (b) Model 2 and (c) Model 3. Green bars represent cases from the improved group, and red bars represent cases from the unimproved group. DNN, deep neural network.

4. Discussion

To our knowledge, this is the first study to use AI to predict the long‐term treatment outcome of TMD patients. In this study, we developed prediction models for TMD prognosis using clinical and haematological features based on machine learning. Despite the relatively small sample size, our models demonstrated strong performance. The decision tree‐based model achieved an accuracy of 0.906 and an F1‐score of 0.800. Similarly, the DNN model, which used fewer features, also performed well with an accuracy of 0.906 and an F1‐score of 0.769. These results suggest that even with a limited sample size, this approach shows promise and could potentially be applied to larger datasets to develop clinically useful prediction models in the future.

In this study, the importance of features was calculated based on the decision tree algorithm to identify key predictors influencing prognosis. Notably, subjective pain‐related features ranked highly among the top 10 most important predictors, which is consistent with previous studies [12, 17]. On the other hand, objective features based on the RDC/TMD, such as diagnostic classifications, capsule and muscle palpation and jaw opening, were not included among the top predictors. Although the RDC/TMD and more recently established DC/TMD have succeeded in providing a standardised platform for the diagnosis of TMD [23], it has been pointed out that axis I diagnoses suffer from poor inter‐examiner reliability and reproducibility depending on examiner ability and diagnosis subtype. Furthermore, the system is not based on differentiating the aetiology of the disorder but rather on grouping according to symptom characteristics based on physical examinations [24]. Previous studies aimed at identifying biomarkers for pain prognosis have reported that subjective pain complaints are more closely related to long‐term outcomes than objective physical examination findings in TMD [25]. This aligns with the changes seen in fibromyalgia diagnostic criteria, which have shifted from relying on objective palpation of tender points to a focus on subjective reports of pain regions [26]. The given results support the need to prioritise and incorporate simple subjective pain assessments as key components in the development of a future aetiology‐based diagnostic system for TMD, which is capable of predicting prognosis.

Among the top 10 features identified in our study, aside from the subjective pain‐related features, the remaining key predictors were haematologic parameters (MCHC, Hgb, Hct, LMR, ESR). All of these features show a certain association with systemic inflammation [27, 28, 29]. The predictive performance of the DNN‐based model improved significantly when these haematologic features were added to the subjective pain‐related features. This suggests that haematologic features reflecting systemic inflammation could be crucial factors in predicting the long‐term outcomes of TMD patients. These findings are consistent with previous research [17], which highlighted the potential role of non‐specific inflammation in the chronicity of TMD.

There is growing evidence that systemic inflammation, including elevated inflammatory biomarkers, such as cytokines, is closely related to various types of chronic pain disorders [30]. Although the exact mechanisms by which systemic inflammation influences chronic pain are yet to be fully understood, inflammation can sensitise nociceptors and affect pain transmission pathways in both the central and peripheral nervous systems through both direct and indirect ways, making individuals more susceptible to prolonged pain [31, 32, 33]. Similar to other chronic pain conditions, there have been reports suggesting that aberrations in systemic inflammation markers are associated with the worsening of symptoms, chronicity and poor treatment outcomes in TMD [15, 16, 20]. Based on these findings, it can be inferred that systemic inflammation plays an important role in the progression and exacerbation of TMD, and an evaluation of systemic inflammation should be an integral part of an aetiology‐based diagnostic system for TMD.

Haematologic markers associated with systemic inflammation used in this study offer a simpler and more cost‐effective means of assessing inflammatory status compared to direct cytokine analysis. In our study, MCHC, Hgb, Hct, LMR and ESR, which are generally included in routine bloodwork or can be derived from results of such simple blood tests, were identified as key predictors of TMD prognosis. Therefore, these systemic biomarkers could serve as valuable tools for predicting long‐term outcomes of TMD, and be included in an aetiology‐based diagnostic system for TMD as alternatives to more complex and costly cytokine analysis.

Several limitations should be acknowledged of this study. First, due to the small sample size, our prediction model, although it showed reasonably good performance, did not reach the level of accuracy required for immediate clinical use. A limited sample size may also increase the risk of overfitting during model training and reduce the generalisability of the results to broader TMD populations. Nonetheless, the promising results suggest the potential for further improvements when larger datasets are utilised and this could allow the application of various classification thresholds for pain improvement including both relative (e.g., 30% or ≤ 50%) and absolute (e.g., 2 or 3 point) criteria. Second, the actual utility of the selected features remains uncertain. Given the small sample size, it is possible that certain data points may have disproportionately influenced the model's predictions. Finally, the AI models employed in this study are inherently black boxes, indicating that we cannot fully understand how the selected features influence the prognosis.

To address the limitations outlined above and further improve the clinical applicability of AI‐based prediction models for TMD, several directions for future research should be considered. First, the generalisability of the current findings should be validated in larger and more diverse patient populations, taking into account variations in age, gender and lifestyle habits based on real‐life data. Second, future studies should aim to integrate various longitudinal data, including long‐term treatment responses and follow‐up outcomes, to capture the dynamic nature of TMD progression and therapeutic response. Third, while this study demonstrated the prognostic potential of haematologic markers related to systemic inflammation, further research is needed to investigate the mechanisms by which such inflammatory processes affect TMD to eventually establish criteria for prediction. Such criteria should involve the timing of early detection and initiation, and selection of treatment methods based on these markers. Finally, the integration of explainable AI approaches could enhance the transparency and interpretability of the model outputs, ultimately supporting more informed clinical decision‐making.

In conclusion, this study demonstrated the potential of machine learning models in predicting TMD prognosis using simple and easily obtainable clinical and haematological features. Our findings also highlighted that subjective pain reports and haematologic markers, which could be related to systemic inflammation, were key components of the long‐term prediction model.

These findings underscore the importance of incorporating these two elements as integral components of an aetiology‐based diagnostic system for TMD to enhance its predictive accuracy and clinical effectiveness.

Author Contributions

Moon Jong Kim: contributed to conception and design, data interpretation, drafted and critically revised the manuscript. Taegun An: contributed to data interpretation, performed all statistical analyses, drafted the manuscript. Il‐San Cho: contributed to the design, data acquisition and critically revised the manuscript. Changhee Joo: contributed to conception and design, data interpretation and critically revised the manuscript. Ji Woon Park: contributed to conception and design, data acquisition and interpretation and critically revised the manuscript. All authors gave their final approval and agree to be accountable for all aspects of the work.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The raw data supporting this study are not in the public domain but are available upon reasonable request from the corresponding author.