Analysis of factors associated with persistent night pain after locking plate fixation of proximal humerus fractures and development of a prediction model

Abstract

Background

Proximal humerus fractures are common fragility fractures in older adults. Open reduction and locking plate internal fixation can restore fracture alignment and facilitate early rehabilitation; however, some patients still develop persistent night pain after surgery, potentially impairing sleep and tolerance of rehabilitation. This study aimed to analyze factors associated with persistent night pain 3 months after locking plate fixation for proximal humerus fractures and to construct a preliminary preoperative risk model and a complete risk stratification model after immediate postoperative radiographic assessment.

Methods

This single-center retrospective observational study included 331 patients with proximal humerus fractures who underwent open reduction and locking plate internal fixation at our hospital between January 1, 2021, and December 31, 2024, and had completed 3-month postoperative follow-up. The primary outcome was persistent night pain at 3 months after surgery. General clinical variables, fracture characteristics, preoperative radiographic parameters, surgical variables, immediate postoperative fixation quality, postoperative recovery, and complications were collected. Eight candidate predictors were prespecified before model development. Variable selection was performed using least absolute shrinkage and selection operator (LASSO) logistic regression with 10-fold cross-validation, and selected variables were refitted using conventional multivariable logistic regression. Model performance was evaluated using receiver operating characteristic (ROC) curves, the area under the curve (AUC), calibration curves, the Hosmer-Lemeshow test, the Brier score, decision curve analysis, 1,000 bootstrap resamples, and internal temporal split validation.

Results

Among the 331 patients, 95 had persistent night pain at 3 months after surgery, corresponding to an incidence of 28.7%. Multivariable logistic regression showed that preinjury sleep disturbance or anxiety/depression, a lower deltoid tuberosity index (DTI), greater displacement of the greater tuberosity, preoperative loss of medial support, lack of effective postoperative medial support, and plate malposition were associated with persistent night pain. The complete model had an AUC of 0.814 (95% confidence interval [CI], 0.763–0.864), a bootstrap-corrected concordance index (C-index) of 0.801, a calibration intercept of -0.012, a calibration slope of 0.946, and a Brier score of 0.154. The preoperative model had an AUC of 0.768 (95% CI, 0.710–0.825). Internal temporal split validation demonstrated stable discrimination and prediction error; in sensitivity analyses excluding major complications, using a stricter alternative outcome, and applying multiple imputation, the effect directions of the principal predictors were consistent with those in the complete-case analysis.

Conclusions

Persistent night pain 3 months after locking plate fixation of proximal humerus fractures is not uncommon. Preinjury sleep disturbance or anxiety/depression, lower DTI, greater-tuberosity displacement, medial support loss, inadequate postoperative medial support reconstruction, and plate malposition were associated with this outcome. The preoperative model may provide preliminary preoperative risk information. The complete model should be applied after immediate postoperative radiographic assessment and may help identify patients at increased risk of persistent night pain and inform targeted postoperative management. Clinical application still requires prospective, multicenter external validation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12891-026-10148-w.

Introduction

Proximal humerus fractures are common osteoporosis-related fractures in older adults. Treatment strategies require consideration of fracture morphology, bone quality, patients’ functional demands, and complication risk [1, 2]. For markedly displaced two-part, three-part, and four-part fractures, open reduction and locking plate internal fixation remains a commonly used treatment option. Its goals are to restore the head-shaft angle, tuberosity position, and medial column support and to create a relatively stable mechanical environment for subsequent rehabilitation [3].

Locking plate fixation cannot completely prevent postoperative pain, restricted motion, or implant-related complications. Fracture type, local bone stock, tuberosity reduction, medial column stability, and implant position may all affect maintenance of reduction and functional recovery [4, 5]. Night pain is a common and easily underestimated symptom after shoulder disorders and surgery for shoulder trauma. Unlike pain during daytime activity, night pain occurs during rest or sleep and is often associated with difficulty falling asleep, sleep interruption, changes in sleeping position, and the need for nocturnal analgesia [6, 7].

Sleep disturbance after shoulder surgery is usually most prominent during the early postoperative period, and some patients still have nocturnal symptoms at 3 months after surgery. This time window may reflect early rehabilitation tolerance and the patient’s perception of recovery [8]. Although 3 months after surgery remains an early follow-up stage, it is usually beyond the peak of the acute postoperative inflammatory response and is a key period of transition from passive motion to active motion and functional training. Pain control and rehabilitation adherence at this stage are closely related to short-term recovery of shoulder function [9].

Therefore, persistent night pain at 3 months after surgery may serve as a clinical intermediate outcome reflecting early symptom burden and rehabilitation tolerance after locking plate fixation; its construct and time horizon differ from those of long-term chronic pain and mid- to long-term functional outcomes [10]. On this basis, we retrospectively included patients with proximal humerus fractures treated with open reduction and locking plate internal fixation, used persistent night pain at 3 months after surgery as the primary outcome, and systematically analyzed general clinical factors, fracture morphology, bone stock, and fixation-quality variables. We also constructed a preliminary preoperative risk model and a complete risk-stratification model for use after immediate postoperative radiographic assessment, and evaluated their internal performance and the robustness of the findings to support postoperative risk stratification and subsequent external validation.

Materials and methods

Study design and participants

This was a single-center retrospective observational study reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) recommendations [11]. Development and validation of the prediction models were reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) recommendations [12].

The study protocol was approved by the Ethics Committee of our hospital, and the requirement for written informed consent was waived. Patients with proximal humerus fractures who underwent open reduction and locking plate internal fixation at our hospital from January 1, 2021, to December 31, 2024 were retrospectively screened.

The inclusion criteria were as follows: (1) age ≥ 18 years; (2) an acute proximal humerus fracture confirmed by imaging; (3) treatment with open reduction and locking plate internal fixation; (4) time from injury to surgery ≤ 21 days; (5) completion of 3-month postoperative follow-up with a clearly adjudicated primary outcome; and (6) clinical, radiographic, and follow-up data available for the primary analysis.

The exclusion criteria were as follows: (1) open fracture, pathological fracture, or chronic fracture; (2) previous ipsilateral shoulder fracture or dislocation, massive rotator cuff injury, frozen shoulder, severe glenohumeral osteoarthritis, or previous shoulder surgery; (3) severe nerve injury, vascular injury, symptomatic cervical radiculopathy, fibromyalgia, long-term chronic opioid use, or other chronic pain syndromes that could substantially affect pain assessment; (4) bilateral proximal humerus fractures or multiple injuries affecting pain assessment; (5) missing key baseline data, radiographic data, or primary outcome data; and (6) loss to follow-up.

Surgical technique and perioperative management

Fractures were classified according to the Neer system, and the surgical plan was developed using preoperative radiographs, axillary or scapular Y views, and three-dimensional computed tomography (CT) reconstruction when necessary [13]. All operations were performed by senior orthopedic trauma surgeons at our hospital. Surgery was performed under brachial plexus block anesthesia, with or without general anesthesia. Regional block techniques, rescue analgesia, and analgesic medications prescribed at discharge were recorded according to the in-hospital stepwise analgesic pathway [14, 15].

During the study period, our center used a standardized perioperative analgesia and rehabilitation pathway. After surgery, all patients received prophylactic antibiotics, thrombosis-risk assessment, sling protection, stepwise analgesia, and standardized rehabilitation education. The rehabilitation protocol followed a standardized progression and was adjusted according to fixation stability, advancing from finger, wrist, and elbow motion to passive shoulder motion, active-assisted motion, and light resistance training [9, 16].

The perioperative analgesic regimen was categorized according to prespecified criteria as a nonsteroidal anti-inflammatory drug (NSAID)-only regimen, a multimodal regimen including short-term opioid rescue analgesia, or a regimen incorporating regional-block analgesia. Records relating to regional blocks included block type, expected duration of analgesia, and transition to the postoperative stepwise regimen. The timing of rehabilitation initiation and adherence were extracted from medical records, rehabilitation records, and follow-up records and interpreted within the evidence framework for treatment and rehabilitation of adult proximal humerus fractures [16].

Outcome definition

The primary outcome was persistent night pain at 3 months after surgery. Night pain was defined as pain in the affected shoulder during nighttime rest or sleep, accompanied by at least one of the following: difficulty falling asleep, awakening because of pain, repeated changes in sleeping position because of pain, or need for additional nocturnal analgesic medication [17].

Persistent night pain was defined as the continued presence of the above manifestations at 3 months after surgery, with episodes occurring on average at least 3 times per week during the 2 weeks before assessment. This frequency threshold was based on the definition of persistent symptom burden as “at least 3 times per week” in adult insomnia-related diagnoses, and the 2-week recall window was used to reduce random misclassification caused by single-day activity load or short-term medication fluctuations [17].

The primary outcome was adjudicated preferentially using standardized outpatient interview records. For patients who did not return to the hospital, researchers who were not involved in surgery, radiographic evaluation, or model construction verified the frequency of night pain, sleep interruption, position changes, and nocturnal rescue analgesia using a standardized telephone script and completed adjudication according to the same criteria. Sleep interruption, position changes, and nocturnal rescue analgesia at 3 months after surgery were used only as components of primary outcome adjudication. They were not defined as separate secondary outcomes, were not included as predictors in the models, and were not interpreted as independent postoperative outcomes in between-group comparisons.

The Pittsburgh Sleep Quality Index (PSQI) was not used as the primary outcome instrument in this study. Although the PSQI permits systematic assessment of sleep quality, original scale scores were not consistently available in the retrospective medical records [18]. Resting pain at 24 h after surgery and pain intensity during follow-up were recorded using a 0–10 visual analog scale (VAS), where 0 indicated no pain and 10 indicated the worst imaginable pain [19].

Variable definitions and data collection

General clinical variables, fracture and radiographic variables, surgical variables, postoperative recovery, and complications were collected. The deltoid tuberosity index (DTI) was measured on standard preoperative anteroposterior shoulder radiographs. At the most proximal level immediately above the deltoid tuberosity where the medial and lateral outer cortical borders of the humeral shaft first became approximately parallel and the medullary borders were clearly visible, total outer cortical width A and inner medullary diameter B were measured perpendicular to the longitudinal axis of the humeral shaft; DTI = A/B [20].

If the standard anteroposterior radiograph was of insufficient quality to identify the cortical borders, DTI was recorded as missing and handled according to the missing-data workflow. Reproducibility of DTI measurements was evaluated using interobserver and intraobserver agreement [21]. Osteoporosis-related history was separated into two descriptive variables: a previous definitive diagnosis of osteoporosis and previous or current anti-osteoporosis treatment. Management of the long head of the biceps tendon was defined as tendon-sheath release or debridement, tenodesis, or tenotomy; simple exposure was not counted for this variable. American Society of Anesthesiologists (ASA) physical status was recorded as a general clinical variable.

Radiographic evaluation

Preoperative medial support loss was defined as preoperative imaging showing discontinuity of the medial cortex, a medial cortical gap or step-off > 2 millimeters (mm), medial metaphyseal comminution length ≥ 10 mm, or a varus head-shaft angle < 120 degrees with medial hinge disruption [22]. Effective postoperative medial support was defined as postoperative imaging showing medial cortical contact with a gap ≤ 2 mm, stable medial impaction, or at least one calcar screw located within 5–10 mm of the inferomedial subchondral bone of the humeral head and supporting the medial column [23].

Appropriate plate position was defined as the superior edge of the plate located 5–8 mm below the summit of the greater tuberosity, lateral prominence ≤ 2 mm, and no risk of subacromial impingement on dynamic fluoroscopy [24]. The anteroposterior position of the plate was determined primarily using the lateral edge of the bicipital groove and the lateral cortex of the greater tuberosity as anatomical references. Screw trajectories were required to avoid penetration of the articular surface and to cover the inferomedial support region of the humeral head as much as possible [25]. Satisfactory reduction was defined as simultaneous achievement of a head-shaft angle of 125–145 degrees, head-shaft translation or step-off ≤ 5 mm, residual greater-tuberosity displacement ≤ 5 mm, humeral head height difference ≤ 5 mm, effective medial support, and absence of intra-articular screw penetration [26].

Loss of reduction, screw cut-out, and fixation failure were determined by comprehensive assessment of immediate postoperative and 3-month follow-up imaging and recorded according to commonly used radiographic complication criteria after locking plate fixation [27]. Avascular necrosis of the humeral head was determined on the basis of sclerosis, collapse, or segmental necrotic changes on follow-up radiographs or CT [28]. All radiographic measurements were performed independently by two physicians experienced in shoulder and elbow trauma imaging, both of whom were blinded to primary outcome status. Images from 50 randomly selected cases were used for repeat measurements. Interobserver agreement for key radiographic and technical variables was an intraclass correlation coefficient (ICC) of 0.88–0.91 or kappa (κ) of 0.79–0.87, and intraobserver agreement was an ICC of 0.92–0.94 or κ of 0.85–0.91 (Supplementary Table 1).

Follow-up protocol and blinded outcome adjudication

Routine outpatient follow-up was conducted at 1 and 3 months after surgery; standardized telephone follow-up was used for patients who did not return as scheduled. The primary outcome was independently adjudicated by two researchers who were not involved in surgery, radiographic measurement, or model development according to prespecified criteria. During adjudication, the assessors were provided only with anonymized follow-up records and standardized interview forms and were blinded to the surgeon, surgical approach, effective postoperative medial support, plate position, reduction quality, and other radiographic technical variables. If the two assessors disagreed, a third researcher made the final adjudication while remaining masked to the same radiographic technical variables. Interobserver agreement between the two primary outcome assessors for persistent night pain at 3 months after surgery was evaluated using Cohen’s κ coefficient and was high (κ = 0.91; 95% confidence interval [CI], 0.86–0.96) (Supplementary Table 1).

Statistical analysis

The distribution of continuous variables was assessed using the Shapiro-Wilk test. Approximately normally distributed continuous variables are presented as mean ± standard deviation and were compared between groups using Welch’s t test. Non-normally distributed variables are presented as median (interquartile range) and were compared using the Mann-Whitney U test. Categorical variables are presented as number (percentage). Pearson’s χ² test was used for multicategory contingency tables. For 2 × 2 contingency tables, the continuity-corrected χ² test was used when all expected cell counts were ≥ 5, and Fisher’s exact test was used when any expected cell count was < 5. All tests were two-sided, and P < 0.05 was considered statistically significant. Univariable comparisons were used only to describe between-group distributions and preliminary associations and were not the sole criterion for inclusion in the model.

All cases that met the eligibility criteria during the study period and had an adjudicable primary outcome entered the candidate cohort. Before modeling, eight candidate predictors for the main model were prespecified and limited according to the prediction time point, number of events, clinical availability, and avoidance of obvious mediators: age, diabetes mellitus, preinjury sleep disturbance or anxiety/depression, DTI, greater-tuberosity displacement distance, preoperative loss of medial support, lack of effective postoperative medial support, and plate malposition. Neer classification, preoperative head-shaft angle, medial column comminution, operative time, bone grafting, reduction quality, and postoperative recovery indicators were used only for description or sensitivity analyses and were not candidate predictors in the main model, thereby reducing the risks of overfitting, variable overlap, and confounding along the causal pathway. With 95 outcome events and 8 candidate predictors, the events per parameter (EPP) before selection was 11.9. The EPP was 15.8 for the 6 predictors in the final complete model and 23.8 for the 4 predictors in the preoperative model.

Least absolute shrinkage and selection operator (LASSO) logistic regression was used, with the penalty parameter corresponding to the one-standard-error rule (λ1se) selected by 10-fold cross-validation. Selected variables were then entered into conventional multivariable logistic regression. Linearity of continuous variables with the logit was evaluated using the Box-Tidwell test and restricted cubic splines (RCS), and multicollinearity was assessed using the variance inflation factor (VIF). Model performance was evaluated using receiver operating characteristic (ROC) curves and the area under the curve (AUC), calibration curves, the Hosmer-Lemeshow test, calibration intercept, calibration slope, Brier score, decision curve analysis (DCA), 1,000 bootstrap resamples, and internal temporal split validation. The 95% confidence intervals for AUCs were estimated using the DeLong method. For internal temporal split validation, cases from 2021 to 2022 formed the training set and cases from 2023 to 2024 formed the test set. The six prespecified predictors in the complete model were refitted in the training set, and AUC and Brier score were evaluated in the test set using frozen coefficients, without variable reselection or coefficient updating in the test set. The sensitivity analysis excluding major complications used the patient as the unit of analysis and excluded any patient who experienced loss of reduction, deep infection, fixation failure, screw cut-out, or reoperation within 3 months after surgery. Complication categories could overlap within a patient, but each patient was excluded only once. In the remaining cohort, multivariable logistic regression was refitted using the six predictors specified in the complete model, without repeating variable selection. Adjusted odds ratios (ORs), 95% confidence intervals, and P values were used to evaluate the direction of associations; AUC and Brier score were reported as the apparent performance of the refitted model in that cohort. This analysis assessed whether the principal associations remained consistent among patients without these complications. The stricter alternative outcome was defined as night pain occurring on average at least 4 times per week during the 2 weeks before the 3-month assessment or continued use of nocturnal rescue analgesia. The primary analysis used complete cases. Missing primary outcomes were not imputed; patients with known outcomes but missing covariates were included in a multiple-imputation sensitivity analysis. Multiple imputation by chained equations included the primary outcome, the eight candidate predictors, and baseline variables associated with missingness and generated 20 imputed datasets. Regression coefficients and standard errors were pooled according to Rubin’s rules, and AUCs were estimated separately in each imputed dataset and summarized by their mean.

Results

Patient screening, general clinical characteristics, and follow-up mode

A total of 412 patients with proximal humerus fractures treated with locking plate fixation were screened. Fifty-four patients were excluded because they did not meet the clinical criteria. Reasons for exclusion were classified mutually exclusively according to the primary reason: open, pathological, or chronic fractures in 9 patients; previous ipsilateral shoulder fracture/dislocation, massive rotator cuff tear, frozen shoulder, severe glenohumeral osteoarthritis, or previous shoulder surgery in 15; severe nerve/vascular injury, symptomatic cervical radiculopathy, fibromyalgia, long-term opioid use, or another chronic pain syndrome that substantially affected pain assessment in 14; bilateral proximal humerus fractures or multiple injuries affecting pain assessment in 16; and an injury-to-surgery interval > 21 days in 0. An additional 7 patients had missing primary outcomes and 20 had known outcomes but missing covariates; 11 patients in the latter group had unmeasurable DTI. Finally, 331 patients entered the complete-case analysis. Persistent night pain at 3 months after surgery occurred in 95 patients (28.7%), whereas 236 did not experience this outcome (Fig. 1). Types and handling of missing data are presented in the Supplementary Materials (Supplementary Table 2).

Fig. 1.

Patient screening flowchart. A total of 412 patients with proximal humerus fractures treated with locking plate fixation were screened; 54 were excluded because they did not meet the clinical criteria, and 358 met the clinical eligibility criteria. Seven patients had missing primary outcomes, and 20 had known outcomes but missing covariates; 11 patients with unmeasurable deltoid tuberosity index (DTI) were included among these 20 patients. Finally, 331 patients were included in the complete-case analysis. At 3 months after surgery, the primary outcome was assessed during outpatient follow-up in 248 patients and by telephone in 83. Persistent night pain occurred in 95 patients (28.7%; outpatient follow-up, n = 69; telephone follow-up, n = 26), and no persistent night pain occurred in 236 patients (71.3%; outpatient follow-up, n = 179; telephone follow-up, n = 57). n, number of patients. Outcome percentages were calculated using the complete-case cohort of 331 patients

Compared with patients without persistent night pain, those with persistent night pain were older, had higher proportions of diabetes mellitus and a previous definitive diagnosis of osteoporosis, showed a shift toward higher ASA physical status categories, and more commonly had preinjury sleep disturbance or anxiety/depression (all P < 0.05). No statistically significant between-group differences were observed in sex, body mass index, side of injury, injury mechanism, smoking history, alcohol history, hypertension, coronary heart disease, previous or current anti-osteoporosis treatment, or long-term glucocorticoid use (all P > 0.05) (Table 1). The distribution of outpatient versus telephone follow-up did not differ significantly between the groups (P = 0.638) (Fig. 1).

Table 1.

Comparison of general clinical characteristics between the two groups

Variable	Persistent night pain group (n = 95)	No persistent night pain group (n = 236)	Statistic	P value
Age (years), mean ± SD	66.7 ± 10.8	62.7 ± 11.4	t = 2.999	0.003
Female, n (%)	71 (74.7)	162 (68.6)	χ² = 0.932	0.334
Body mass index (kg/m²), mean ± SD	24.3 ± 3.4	24.1 ± 3.2	t = 0.492	0.623
Right-sided injury, n (%)	46 (48.4)	124 (52.5)	χ² = 0.310	0.577
High-energy injury, n (%)	35 (36.8)	62 (26.3)	χ² = 3.161	0.075
Smoking history, n (%)	31 (32.6)	64 (27.1)	χ² = 0.755	0.385
Alcohol history, n (%)	24 (25.3)	50 (21.2)	χ² = 0.435	0.510
Diabetes mellitus, n (%)	28 (29.5)	37 (15.7)	χ² = 7.318	0.007
Hypertension, n (%)	44 (46.3)	85 (36.0)	χ² = 2.603	0.107
Coronary heart disease, n (%)	16 (16.8)	25 (10.6)	χ² = 1.895	0.169
Previous definitive diagnosis of osteoporosis, n (%)	33 (34.7)	52 (22.0)	χ² = 5.081	0.024
Previous or current anti-osteoporosis treatment, n (%)	24 (25.3)	38 (16.1)	χ² = 3.157	0.076
Long-term glucocorticoid use, n (%)	11 (11.6)	13 (5.5)	χ² = 2.864	0.091
ASA physical status I/II/III, n	13/56/26	58/145/33	χ² = 10.624	0.005
Preinjury sleep disturbance or anxiety/depression, n (%)	26 (27.4)	24 (10.2)	χ² = 14.312	< 0.001

Continuous variables are presented as mean ± standard deviation and categorical variables as n (%)

SD standard deviation, n number of patients, % percentage, kg/m² kilograms per square meter, t statistic from Welch’s t test, χ² Pearson’s χ² test for multicategory variables and the continuity-corrected χ² test for 2 × 2 contingency tables when all expected cell counts were ≥ 5, P value two-sided test probability, ASA American Society of Anesthesiologists physical status classification

Fracture characteristics, preoperative radiographic variables, and surgical variables

The distribution of Neer classification in the persistent night pain group shifted toward more complex fractures. Greater-tuberosity displacement, medial column comminution, and medial hinge disruption were more common, and the preoperative head-shaft angle was smaller, the humeral head height difference and greater-tuberosity displacement distance were greater, and DTI was lower (all P < 0.05). No statistically significant between-group differences were observed in head-splitting fracture, dislocation, or fracture-line extension into the humeral shaft (all P > 0.05) (Table 2).

Table 2.

Comparison of fracture characteristics and preoperative radiographic variables between the two groups

Variable	Persistent night pain group (n = 95)	No persistent night pain group (n = 236)	Statistic	P value
Neer two-/three-/four-part classification, n	10/46/39	51/117/68	χ² = 7.672	0.022
Greater-tuberosity displacement, n (%)	53 (55.8)	91 (38.6)	χ² = 7.496	0.006
Medial column comminution, n (%)	51 (53.7)	81 (34.3)	χ² = 9.799	0.002
Head-splitting fracture, n (%)	13 (13.7)	16 (6.8)	χ² = 3.222	0.073
Dislocation, n (%)	11 (11.6)	12 (5.1)	χ² = 3.471	0.062
Fracture-line extension into the humeral shaft, n (%)	18 (18.9)	30 (12.7)	χ² = 1.651	0.199
Preoperative head-shaft angle (°), mean ± SD	114.8 ± 12.5	121.5 ± 12.4	t = -4.421	< 0.001
Humeral head height difference (mm), mean ± SD	9.6 ± 3.7	7.3 ± 3.4	t = 5.234	< 0.001
Greater-tuberosity displacement distance (mm), mean ± SD	8.7 ± 4.1	5.9 ± 3.8	t = 5.738	< 0.001
Medial hinge disruption, n (%)	57 (60.0)	84 (35.6)	χ² = 15.518	< 0.001
DTI, mean ± SD	1.35 ± 0.16	1.46 ± 0.18	t = -5.454	< 0.001

Continuous variables are presented as mean ± standard deviation and categorical variables as n (%)

SD standard deviation, n number of patients, % percentage, mm millimeter, ° degree, t statistic from Welch’s t test, χ² Pearson’s χ² test for multicategory variables and the continuity-corrected χ² test for 2 × 2 contingency tables when all expected cell counts were ≥ 5, P value two-sided test probability, DTI deltoid tuberosity index

The persistent night pain group had a longer operative time, greater estimated intraoperative blood loss, more intraoperative fluoroscopy exposures, a smaller postoperative head-shaft angle, and a greater postoperative residual humeral head height difference. Preoperative loss of medial support, a suboptimal proximal screw tip-articular surface relationship, lack of effective postoperative medial support, use of bone graft or bone substitute material, plate malposition, management of the long head of the biceps tendon, and unsatisfactory reduction were more common in this group (all P < 0.05). No statistically significant between-group differences were observed in time from injury to surgery, surgical approach, number of proximal screws, or perioperative analgesic regimen (all P > 0.05) (Table 3).

Table 3.

Comparison of surgical variables, immediate postoperative fixation quality, and perioperative analgesia between the two groups

Variable	Persistent night pain group (n = 95)	No persistent night pain group (n = 236)	Statistic	P value
Time from injury to surgery (d), mean ± SD	6.2 ± 3.4	5.5 ± 3.0	t = 1.751	0.082
Surgical approach: deltopectoral/deltoid-splitting, n	60/35	129/107	χ² = 1.665	0.197
Operative time (min), mean ± SD	112.8 ± 24.6	101.7 ± 22.8	t = 3.791	< 0.001
Estimated intraoperative blood loss (mL), mean ± SD	208.0 ± 88.0	174.0 ± 76.0	t = 3.303	0.001
Intraoperative fluoroscopy exposures, mean ± SD	13.6 ± 3.9	12.1 ± 3.5	t = 3.258	0.001
Preoperative loss of medial support, n (%)	58 (61.1)	80 (33.9)	χ² = 19.443	< 0.001
Postoperative head-shaft angle (°), mean ± SD	128.9 ± 7.5	133.4 ± 6.9	t = -5.051	< 0.001
Postoperative residual humeral head height difference (mm), mean ± SD	6.1 ± 2.8	4.6 ± 2.4	t = 4.587	< 0.001
Suboptimal proximal screw tip-articular surface relationship, n (%)	28 (29.5)	42 (17.8)	χ² = 4.860	0.027
Lack of effective postoperative medial support, n (%)	40 (42.1)	46 (19.5)	χ² = 16.855	< 0.001
Bone graft or bone substitute material, n (%)	36 (37.9)	52 (22.0)	χ² = 7.937	0.005
Number of proximal screws, mean ± SD	6.6 ± 1.0	6.8 ± 0.9	t = -1.693	0.092
Plate malposition, n (%)	26 (27.4)	26 (11.0)	χ² = 12.469	< 0.001
Management of the long head of the biceps tendon, n (%)	30 (31.6)	43 (18.2)	χ² = 6.276	0.012
Unsatisfactory reduction, n (%)	42 (44.2)	44 (18.6)	χ² = 21.712	< 0.001
Perioperative analgesic regimen: NSAID/opioid multimodal/regional block, n	28/47/20	58/132/46	χ² = 1.231	0.540

Continuous variables are presented as mean ± standard deviation and categorical variables as n (%)

SD standard deviation, n number of patients, % percentage, d day, min minute, mL milliliter, mm millimeter, ° degree, NSAID nonsteroidal anti-inflammatory drug, t statistic from Welch’s t test, χ² Pearson’s χ² test for multicategory variables and the continuity-corrected χ² test for 2 × 2 contingency tables when all expected cell counts were ≥ 5, P value two-sided test probability

Postoperative recovery and complications

The persistent night pain group had a longer hospital stay, a higher resting VAS score at 24 h after surgery, and later initiation of passive and active motion. At 3 months after surgery, daytime resting pain, pain during activity, shoulder stiffness, loss of reduction, deep infection, fixation failure, screw cut-out, and reoperation were more common, whereas forward flexion, abduction, external rotation at the side, and the radiographic fracture-healing rate were lower (all P < 0.05). Additional analgesia within 24 h after surgery did not differ significantly between the groups (P > 0.05) (Table 4).

Table 4.

Comparison of postoperative recovery and complications between the two groups

Variable	Persistent night pain group (n = 95)	No persistent night pain group (n = 236)	Statistic	P value
Hospital stay (d), mean ± SD	9.2 ± 3.1	8.0 ± 2.7	t = 3.302	0.001
Resting VAS score at 24 h after surgery (points), mean ± SD	4.8 ± 1.3	3.9 ± 1.2	t = 5.823	< 0.001
Additional analgesia within 24 h after surgery, n (%)	38 (40.0)	72 (30.5)	χ² = 2.339	0.126
Time to initiation of passive motion (d), mean ± SD	4.6 ± 1.7	3.8 ± 1.5	t = 4.002	< 0.001
Time to initiation of active motion (d), mean ± SD	32.6 ± 8.9	28.1 ± 7.7	t = 4.320	< 0.001
Daytime resting pain at 3 months after surgery, n (%)	37 (38.9)	21 (8.9)	χ² = 40.265	< 0.001
Pain during activity at 3 months after surgery, n (%)	76 (80.0)	104 (44.1)	χ² = 33.818	< 0.001
Forward flexion at 3 months after surgery (°), mean ± SD	101.2 ± 22.8	124.5 ± 20.7	t = -8.631	< 0.001
Abduction at 3 months after surgery (°), mean ± SD	92.7 ± 21.4	115.7 ± 20.2	t = -8.987	< 0.001
External rotation at the side at 3 months after surgery (°), mean ± SD	24.3 ± 8.9	35.1 ± 9.7	t = -9.728	< 0.001
Radiographic fracture healing at 3 months after surgery, n (%)	60 (63.2)	191 (80.9)	χ² = 10.726	0.001
Shoulder stiffness, n (%)	49 (51.6)	53 (22.5)	χ² = 25.595	< 0.001
Loss of reduction, n (%)	14 (14.7)	11 (4.7)	χ² = 8.458	0.004
Deep infection, n (%)	6 (6.3)	4 (1.7)	Fisher’s exact test	0.036
Fixation failure, n (%)	8 (8.4)	5 (2.1)	Fisher’s exact test	0.012
Screw cut-out, n (%)	11 (11.6)	6 (2.5)	Fisher’s exact test	0.002
Reoperation, n (%)	7 (7.4)	5 (2.1)	Fisher’s exact test	0.044

Continuous variables are presented as mean ± standard deviation and categorical variables as n (%)

SD standard deviation, n number of patients, % percentage, d day, h hour, ° degree, VAS visual analog scale, t statistic from Welch’s t test, χ² continuity-corrected χ² test for 2 × 2 contingency tables when all expected cell counts were ≥ 5, Fisher Fisher’s exact test when any expected cell count was < 5, P value two-sided test probability

Sleep interruption, changes in sleeping position, and nocturnal rescue analgesia at 3 months after surgery were components of the primary outcome and were not treated as independent secondary outcomes or candidate predictors. The trajectory of night pain at 1 month after surgery was used only in a descriptive follow-up analysis.

Model construction, model performance, and sensitivity analyses

After LASSO selection, the complete model retained preinjury sleep disturbance or anxiety/depression, DTI, greater-tuberosity displacement distance, preoperative loss of medial support, lack of effective postoperative medial support, and plate malposition. VIF values ranged from 1.11 to 2.34, indicating no obvious multicollinearity. All four predictors in the preoperative model and all six predictors in the complete model were associated with persistent night pain: preinjury sleep disturbance or anxiety/depression, greater-tuberosity displacement distance, preoperative loss of medial support, lack of effective postoperative medial support, and plate malposition were associated with higher odds of the outcome, whereas a higher DTI was associated with lower odds (all P < 0.05). Both models included a linear predictor with an intercept. Let the linear predictor (LP) be defined by the corresponding equation in Table 5; the individual predicted probability is calculated as P = 1/[1 + exp(-LP)]. The complete model had a higher AUC than the preoperative model. Its bootstrap-corrected concordance index (C-index) decreased only slightly, the calibration intercept was close to 0, the calibration slope was close to 1, the Brier score was lower than that of the preoperative model, and the Hosmer-Lemeshow test was not statistically significant (P > 0.05) (Table 5; Figs. 2, 3, 4 and 5).

Table 5.

Equations and model performance of the preoperative and complete models

Model/variable	β coefficient or metric	OR (95% CI)	P value	Coding description
Preoperative model intercept	-1.280	Not applicable	Not applicable	Constant term in the centered formula
Preoperative model: preinjury sleep disturbance or anxiety/depression	0.812	2.25 (1.23–4.12)	0.008	Yes = 1; no = 0
Preoperative model: DTI_c	-0.187	0.83 (0.75–0.93) per 0.1 increase	0.001	DTI_c = (DTI − 1.40)/0.10
Preoperative model: GT_c	0.085	1.09 (1.04–1.13) per 1-mm increase	< 0.001	GT_c = greater-tuberosity displacement distance − 6.70
Preoperative model: preoperative loss of medial support	0.724	2.06 (1.21–3.52)	0.008	Yes = 1; no = 0
Complete model intercept	-1.627	Not applicable	Not applicable	Constant term in the centered formula
Complete model: preinjury sleep disturbance or anxiety/depression	0.859	2.36 (1.27–4.39)	0.007	Yes = 1; no = 0
Complete model: DTI_c	-0.198	0.82 (0.73–0.92) per 0.1 increase	0.001	DTI_c = (DTI − 1.40)/0.10
Complete model: GT_c	0.086	1.09 (1.04–1.14) per 1-mm increase	< 0.001	GT_c = greater-tuberosity displacement distance − 6.70
Complete model: preoperative loss of medial support	0.683	1.98 (1.10–3.57)	0.023	Yes = 1; no = 0
Complete model: lack of effective postoperative medial support	0.932	2.54 (1.37–4.71)	0.003	Yes = 1; no = 0
Complete model: plate malposition	0.779	2.18 (1.08–4.40)	0.029	Yes = 1; no = 0
Preoperative model performance	AUC = 0.768 (95% CI, 0.710–0.825)	Bootstrap C = 0.751; Brier = 0.172	—	Calibration intercept = -0.020; calibration slope = 0.925
Complete model performance	AUC = 0.814 (95% CI, 0.763–0.864)	Bootstrap C = 0.801; Brier = 0.154	HL P = 0.661	Calibration intercept = -0.012; calibration slope = 0.946

All binary variables were coded as yes = 1 and no = 0. Preoperative model: LP = -1.280 + 0.812 × preinjury sleep disturbance or anxiety/depression − 0.187 × DTI_c + 0.085 × GT_c + 0.724 × preoperative loss of medial support. Complete model: LP = -1.627 + 0.859 × preinjury sleep disturbance or anxiety/depression − 0.198 × DTI_c + 0.086 × GT_c + 0.683 × preoperative loss of medial support + 0.932 × lack of effective postoperative medial support + 0.779 × plate malposition. Individual predicted probability: P = 1/[1 + exp(-LP)]

β logistic regression coefficient, OR odds ratio, CI confidence interval, P value two-sided test probability, DTI deltoid tuberosity index, DTI_c DTI centered at 1.40 and scaled by 0.10, GT_c greater-tuberosity displacement distance centered at 6.70 mm, mm millimeter, AUC area under the receiver operating characteristic curve, C-index concordance index, bootstrap C bootstrap-corrected C-index, Brier Brier score, HL Hosmer-Lemeshow test, LP linear predictor

Fig. 2.

Assessment of logit linearity for continuous variables. The complete-case cohort included 331 patients, of whom 95 had persistent night pain at 3 months after surgery. Panel A shows the relationship between the deltoid tuberosity index (DTI) and the predicted logit, and Panel B shows the relationship between greater-tuberosity displacement distance and the predicted logit. Solid lines represent linear terms, dashed lines represent restricted cubic spline (RCS) fits, and shaded areas represent 95% confidence intervals (CIs). For DTI, the Box-Tidwell test yielded P = 0.684 and the nonlinearity test yielded P = 0.571. For greater-tuberosity displacement distance, the Box-Tidwell test yielded P = 0.293 and the nonlinearity test yielded P = 0.318. No significant nonlinearity was identified, and both variables were therefore entered into the models as continuous variables

Fig. 3.

Nomograms of the preoperative and complete models. The complete-case cohort included 331 patients, of whom 95 experienced the primary outcome. Panel A shows the preoperative model, which includes preinjury sleep disturbance or anxiety/depression, the deltoid tuberosity index (DTI), greater-tuberosity displacement distance, and preoperative loss of medial support. Panel B shows the complete model, which additionally includes lack of effective postoperative medial support and plate malposition after immediate postoperative radiographic assessment. Continuous predictors are displayed in their original units, whereas model calculations use the centered variables and regression coefficients reported in Table 5. Binary variables were coded as yes = 1 and no = 0. Let the linear predictor be denoted by LP. The individual predicted probability is calculated as P = 1/[1 + exp(-LP)]. GT, greater tuberosity; mm, millimeter

Fig. 4.

Receiver operating characteristic (ROC) curves for the preoperative and complete models. The complete-case cohort included 331 patients, of whom 95 experienced the primary outcome. The x-axis indicates 1 - specificity, the y-axis indicates sensitivity, and the gray dashed line is the reference line for no discriminative ability. The preoperative model had an area under the curve (AUC) of 0.768 (95% confidence interval [CI], 0.710–0.825), and the complete model had an AUC of 0.814 (95% CI, 0.763–0.864), indicating improved discrimination after inclusion of immediate postoperative fixation-quality variables. AUC, area under the curve; CI, confidence interval

Fig. 5.

Calibration curve of the complete model. The complete-case cohort included 331 patients, of whom 95 experienced the primary outcome. The x-axis represents the predicted probability from the complete model and the y-axis represents the observed outcome probability; the gray dashed line indicates ideal calibration, the blue solid line indicates the calibration curve of the complete model, and the shaded area indicates the 95% confidence interval (CI). After 1,000 bootstrap resamples, the complete model had a calibration intercept of -0.012, a calibration slope of 0.946, a Brier score of 0.154, and a bootstrap-corrected concordance index (C-index) of 0.801

The complete model had an AUC of 0.814 (95% CI, 0.763–0.864), a bootstrap-corrected C-index of 0.801, a calibration intercept of -0.012, a calibration slope of 0.946, and a Brier score of 0.154; the preoperative model had an AUC of 0.768 (95% CI, 0.710–0.825). Decision curve analysis showed that the complete model provided net benefit within the prespecified threshold range (Fig. 6). In the sensitivity analysis excluding major complications, the five complication categories accounted for 77 occurrences; categories could overlap, and 45 unique patients were excluded after patient-level deduplication, including 27 in the persistent night pain group and 18 in the no-persistent-night-pain group. The final sensitivity-analysis cohort included 286 patients, comprising 68 outcome events and 218 non-events. After refitting the six predictors of the complete model, the effect direction of each predictor was consistent with that in the complete-case analysis (all P < 0.05); the apparent AUC was 0.785 (95% CI, 0.726–0.844), and the apparent Brier score was 0.166. Internal temporal split validation showed stable discrimination and prediction error in the training and test sets. In analyses using the stricter alternative outcome and multiple imputation, the effect directions of the principal predictors were consistent with those in the complete-case analysis (Fig. 7; Supplementary Tables 3–5).

Fig. 6.

Decision curve analysis (DCA). The complete-case cohort included 331 patients, of whom 95 experienced the primary outcome. The x-axis indicates the threshold probability for triggering high-risk management, and the y-axis indicates net benefit. The red curve represents the complete model, the blue dashed curve represents the treat-all strategy, the gray dotted line represents the treat-none strategy, and the light-yellow area represents the prespecified range for interpreting net benefit. Across threshold probabilities of 0.10–0.68, the complete model provided greater net benefit than the treat-all or treat-none strategy. High-risk management included intensified analgesic assessment, sleep and psychological screening, rehabilitation-adherence management, and close radiographic follow-up

Fig. 7.

Internal temporal split validation and sensitivity analyses. The training set included cases from 2021–2022 (166 patients, 45 events), and the test set included cases from 2023–2024 (165 patients, 50 events). Panel A summarizes the areas under the receiver operating characteristic curves (AUCs) of the preoperative model, complete model, training set, test set, stricter alternative outcome, and multiple-imputation analysis. Panel B summarizes the Brier scores of the preoperative model, complete model, training set, test set, and analysis excluding major complications. Panel C shows the adjusted odds ratios (ORs) and 95% confidence intervals (CIs) obtained by refitting the six predictors of the complete model after major complications were excluded. The five major-complication categories could overlap within the same patient; after patient-level deduplication, 45 patients were excluded. The sensitivity analysis included 286 patients, with 68 outcome events and 218 non-events, and yielded an apparent AUC of 0.785 (95% CI, 0.726–0.844) and an apparent Brier score of 0.166. Internal temporal split validation indicated overall stability of model performance, and the analyses excluding major complications, using the stricter alternative outcome, and applying multiple imputation were consistent with the main analysis. DTI, deltoid tuberosity index; GT, greater tuberosity

Discussion

This study showed that the incidence of persistent night pain at 3 months after locking plate fixation of proximal humerus fractures was 28.7%. This outcome was associated with preinjury sleep disturbance or anxiety/depression, lower DTI, a larger displacement distance of the greater tuberosity, preoperative medial support loss, lack of effective postoperative medial support, and plate malposition. The complete model showed good discrimination and calibration and is intended for early risk stratification after immediate postoperative radiographic assessment.

Preinjury sleep disturbance or anxiety/depression was associated with persistent night pain. This finding suggests that baseline sleep and psychological status may influence patients’ perception, reporting, and coping with postoperative pain. Sleep disturbance and postoperative pain behavior have a bidirectional relationship. Early identification of patients with a greater sleep and psychological burden may help refine perioperative management strategies [29].

Lower DTI was associated with persistent night pain and may reflect insufficient local bone stock. Reduced local bone quality may increase the risks of difficulty maintaining reduction, insufficient screw support, and micromotion, thereby affecting the pain experience during early rehabilitation [30].

A larger displacement distance of the greater tuberosity was associated with persistent night pain. Displacement of the greater tuberosity may alter the rotator-cuff lever arm, subacromial space, and tuberosity-healing environment. It may also increase the pain burden during abduction and external-rotation training. Greater-tuberosity fragment morphology and residual displacement are associated with shoulder function and complication risk; therefore, this variable has clear clinical interpretability [31, 32].

Both preoperative medial support loss and lack of effective postoperative medial support were associated with this outcome. The former reflects initial fracture complexity, whereas the latter reflects the mechanical environment after fixation. The two variables may lie on the same clinical pathway; therefore, including both in the model was intended for risk stratification rather than causal decomposition. Studies on prediction of screw perforation and reduction loss after locking plate fixation also suggest that insufficient medial metaphyseal support provides important risk information for fixation-failure-related events [33, 34].

Plate malposition was associated with persistent night pain. A possible explanation is that a plate placed too high or with lateral prominence may reduce the subacromial space and increase the risk of local mechanical irritation. The calcar screw position, plate height, and proximal fixation mode may all affect fixation stability and the risk of subacromial irritation [35, 36]. This variable is a modifiable surgery-related factor, but this study cannot prove that plate malposition directly causes night pain.

Bone grafting, bone substitute material, or other augmentation strategies were more common in the persistent night pain group, suggesting that fracture complexity and poor bone stock may have been more prominent in this group. Augmentation strategies may improve the fixation environment in some osteoporotic or medially unstable cases, but evidence regarding indications, material selection, and long-term benefits remains heterogeneous [37, 38].

Postoperative recovery and complication indicators showed that the persistent night pain group had poorer early functional recovery and more frequent major complications. After excluding, at the patient level, any patient with loss of reduction, deep infection, fixation failure, screw cut-out, or reoperation, the directions of association for the six predictors in the complete model remained consistent with those in the main analysis, indicating that these associations remained consistent among patients without major complications. Complications and reinterventions after locking plate fixation remain important factors affecting the effectiveness of fixation for proximal humerus fractures [39, 40].

This study reported the preoperative and complete models separately. The preoperative model may provide preliminary risk information, whereas the complete model incorporates fixation-quality variables available on immediate postoperative imaging and is therefore more suitable for early risk stratification after surgery. The decision curve showed that the model had net benefit across a threshold probability range of 0.10–0.68. However, decision curves should be linked to explicit clinical actions, such as intensified analgesic review, sleep and psychological screening, rehabilitation-adherence management, and closer imaging follow-up [41].

Model calibration showed a calibration intercept close to 0 and a calibration slope close to 1, and the bootstrap-corrected C-index showed only a limited decrease. Nevertheless, single-center internal validation cannot replace external validation. Before clinical implementation, prediction models still need evaluation of discrimination, calibration, and clinical net benefit across different centers, surgical teams, and follow-up processes [42].

This study has several limitations. First, the single-center retrospective design could not completely eliminate selection bias, information bias, or residual confounding. Second, the retrospective data did not systematically include original scores from validated instruments such as the Pittsburgh Sleep Quality Index (PSQI), the Patient-Reported Outcomes Measurement Information System (PROMIS) Sleep Disturbance scale, the sleep-interference item of the Brief Pain Inventory (BPI), or other validated scales; therefore, the primary outcome may have been subject to misclassification and construct overlap. Third, although the primary outcome was adjudicated independently with masking of key surgical and radiographic technical variables, the original follow-up records arose from routine clinical practice and may still have been subject to uneven information capture. Fourth, factors such as partial rotator cuff injury, cervicogenic pain, preoperative pain intensity, preoperative analgesic use, rehabilitation adherence, and tuberosity-healing quality could not be fully measured. Fifth, internal temporal split validation cannot replace multicenter external validation.

Conclusions

Persistent night pain at 3 months after locking plate fixation of proximal humerus fractures was associated with preinjury sleep disturbance or anxiety/depression, lower DTI, greater-tuberosity displacement, medial support loss, inadequate postoperative medial support reconstruction, and plate malposition. The preoperative model may provide preliminary preoperative risk information. The complete model should be applied after immediate postoperative radiographic assessment and may assist early postoperative risk stratification and highlight modifiable surgery-related factors that warrant attention. This model requires prospective, multicenter external validation before it can be used for routine clinical decision-making.

Authors’ contributions

ZW conceived and designed the study, supervised the overall conduct of the research, interpreted the data, and critically revised the manuscript for important intellectual content. JY participated in study design, collected the clinical and imaging data, verified the data, performed the statistical analysis, organized the results, and drafted the manuscript. As the corresponding author, JY was also responsible for manuscript submission, revision, and communication with the journal. FS, LQ, KL, and HY participated in case data collection, imaging data extraction, and data checking. All authors read and approved the final manuscript.

Funding

Not applicable.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study was reviewed and approved by the Ethics Committee of Fengjie County People’s Hospital (approval number: AF/SS-02/02.0). All procedures performed in this study involving human participants were in accordance with the ethical standards of the institutional research committee and with the Declaration of Helsinki. Because this study was retrospective in nature and used de-identified clinical and imaging data without involving the disclosure of patient privacy, the requirement for written informed consent was waived by the ethics committee.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.