Predicting poor wrist function recovery after internal fixation of distal radius fractures: A retrospective study based on multivariate analysis and machine learning models

Abstract

This study aimed to identify independent risk factors associated with poor wrist function recovery 6 months after internal fixation of distal radius fractures (DRF) and develop machine learning models for early risk prediction. This retrospective study included 328 patients who underwent internal fixation for DRF at our hospital between January 2022 and June 2025. Wrist function was evaluated using the Cooney Wrist Score at 6 months postoperatively. Poor recovery was defined as a Cooney score <75. The overall incidence of poor wrist function recovery was 21.56% (71/328). Clinical variables including demographic characteristics, comorbidities, laboratory indicators, fracture type, postoperative pain scores, and rehabilitation compliance were collected. Rehabilitation compliance referred to patients’ adherence to prescribed postoperative functional training and follow-up plans. Variables with P <.05 in univariate analysis were entered into multivariate logistic regression to identify independent risk factors. These variables were subsequently used to construct random forest (RF), XGBoost, and backpropagation (BP) neural network models. Model performance was evaluated using ROC curves and classification metrics. Multivariate analysis identified low rehabilitation compliance, age ≥65 years, AO type C fractures, postoperative visual analogue scale scores >3 at 1 week, and decreased serum albumin levels as independent predictors of poor wrist function recovery. Among the machine learning models, the RF model demonstrated the best performance (AUC = 0.939, accuracy = 0.946), followed by the BP neural network (AUC = 0.924). The XGBoost model showed comparatively lower predictive performance (AUC = 0.876). Poor wrist function recovery after internal fixation of DRF is influenced by multiple clinical factors. Machine learning models based on identified independent risk factors, particularly the RF model, show strong predictive ability and may assist clinicians in early risk stratification and individualized rehabilitation planning.

1. Introduction

Distal radius fractures (DRF) are a common clinical fracture type, particularly prevalent among the elderly population.^[1] With the accelerating pace of population aging, the incidence of DRF has been increasing annually.^[2] This fracture not only impacts patients’ daily living activities, particularly wrist mobility, but may also lead to long-term functional impairment and reduced quality of life.^[3] Internal fixation surgery, as the primary treatment for DRF, effectively restores fracture stability and provides a foundation for postoperative functional recovery. However, despite surgical resolution of the fracture, significant individual variation exists in wrist function recovery, influenced by complex and diverse factors.^[4]

Numerous studies^[5,6] have examined factors influencing postoperative wrist function recovery in DRF, including age, fracture type, postoperative pain, patient compliance, and comorbidities. While certain factors are recognized as significant predictors of recovery, a systematic approach to accurately identify patients at risk for poor outcomes remains elusive. This lack of prediction poses a major clinical challenge in early identification of high-risk patients for targeted interventions.

With the rapid advancement of artificial intelligence technology, machine learning has demonstrated superior advantages over traditional statistical methods in fracture prognosis prediction, risk factor exploration, and disease risk stratification.^[7–9] Machine learning models can effectively process high-dimensional and multi-category clinical data while automatically identifying latent nonlinear relationships, offering new opportunities for constructing risk prediction tools for poor wrist function recovery. However, research on machine learning for predicting the risk of poor wrist function recovery in patients after internal fixation of DRF remains relatively scarce, and no systematic, mature prediction strategy has yet been established.

To address these limitations, the present study aimed to identify independent risk factors associated with poor wrist function recovery following internal fixation of DRF and to develop predictive models based on machine learning approaches. We hypothesized that integrating multivariate regression analysis with machine learning algorithms would improve risk stratification and enable more accurate early identification of high-risk patients. By establishing and comparing different predictive models, this study seeks to provide a reliable and clinically applicable tool to support individualized rehabilitation planning and postoperative management.

2. Data and methods

2.1. Data sources and collection

This retrospective study collected clinical data from patients hospitalized for DRF between January 2022 and June 2025. Based on our hospital’s patient records, the research aimed to explore factors associated with wrist function recovery following internal fixation surgery for DRF. As a retrospective study, this research complies with the relevant regulations and requirements of the hospital ethics committee, ensuring strict adherence to ethical standards throughout the process. All patient data were anonymized, and the study content did not involve any aspects that could potentially infringe upon patient privacy. Through this retrospective analysis, we aim to comprehensively evaluate key factors influencing postoperative recovery and provide a more scientific basis for future clinical treatment.

2.2. Inclusion and exclusion criteria

Inclusion criteria were: adult patients aged 18 years or older diagnosed with DRF who underwent internal fixation surgery. Subjects must have had complete follow-up records and functional assessment data. Exclusion criteria included: patients with severe underlying conditions (e.g., cardiovascular/cerebrovascular disease, severe hepatic/renal impairment), multiple fractures or fractures with severe complications (e.g., nerve injury, soft tissue damage), postoperative complications (e.g., surgical site infection, nonunion), mental disorders or cognitive impairments preventing treatment compliance or assessment, and failure to complete required postoperative follow-up or functional assessments. Additionally, pregnant or lactating women were excluded.

2.3. Collection of relevant variables

This study categorized relevant information for patients undergoing internal fixation of DRF into 2 groups, specifically including: gender (male/female), age (65 years and older/under 65 years), body mass index (normal BMI <24/overweight or obese BMI ≥24), smoking status (smoker/nonsmoker), drinking habits (frequent drinker/infrequent drinker), hypertension (present/absent), diabetes (present/absent), postoperative 1-week pain score (mild pain VAS score ≤3/severe pain VAS score >3), admission hemoglobin level (normal/decreased), admission albumin level (normal/decreased), patient compliance (high compliance/low compliance), and fracture type (Type A, B, and C).

2.4. Definition of outcomes

For functional recovery assessment in DRF, the Cooney Wrist Score^[10] was employed. This scoring system comprehensively evaluates wrist function recovery by assessing range of motion, strength, functional limitations, and pain. Scores range from 0 to 100, with higher scores indicating better functional restoration. Based on Cooney scores, patients were categorized into 2 groups: – good recovery (Cooney score ≥75): Indicates near- or full-normal wrist function. – poor recovery (Cooney score <75): Indicates incomplete functional recovery, potentially with significant movement limitations or persistent pain. The threshold of 75 points was selected because multiple clinical studies have identified this score as a critical turning point in functional recovery. Scores below 75 typically indicate significant functional impairment and pain, preventing patients from returning to normal daily activities. The outcome assessment time was defined as 6 months post-surgery, when most patients’ functional recovery stabilizes, providing a comprehensive reflection of postoperative recovery. This binary classification standard simplifies rehabilitation assessment for patients and provides clear grouping criteria for subsequent statistical analysis and predictive modeling, thereby offering effective support for clinical treatment and rehabilitation management.

2.5. Statistical analysis

All statistical analyses and model construction were performed using R software (version 4.2.1) and Python (version 3.9). Data preprocessing included verification of data integrity, handling of missing values, categorical variable encoding, and identification of extreme outliers. Continuous variables were assessed for plausibility prior to modeling.

The dataset was randomly divided into a training set (70%) and a test set (30%) using stratified sampling to preserve the outcome distribution. The training set was used for variable selection, model development, and parameter tuning, while the test set was reserved for independent evaluation of model performance.

Given the retrospective design of this study, no formal a priori power calculation was conducted. The sample size was determined by the number of eligible patients during the study period. To ensure model stability and avoid overfitting in multivariable logistic regression, we adhered to the commonly accepted rule of at least 10 events per variable. With 71 cases of poor wrist function recovery, the data supported inclusion of up to approximately 7 independent variables. The final model retained 5 predictors, satisfying events per variable requirements.

Twelve candidate variables were prespecified based on clinical relevance, previously published evidence, and expert consensus. These included demographic characteristics, comorbidities, laboratory indicators, and fracture-related variables commonly associated with functional outcomes after DRF. Univariate analysis (χ² test for categorical variables) was performed within the training set as a screening step. Variables with P <.05 were entered into a multivariable logistic regression model to identify independent predictors.

Significant independent predictors were subsequently used as input features to construct 3 machine learning models: random forest (RF), extreme gradient boosting (XGBoost), and backpropagation neural network (BP). Hyperparameters were optimized using cross-validation within the training set. Continuous variables were standardized when required (e.g., for BP neural network), with transformations performed exclusively on the training data and applied to the test set to prevent data leakage. Early stopping strategies were implemented to further reduce overfitting risk.

Model performance was evaluated on the independent test set using the following metrics: area under the receiver operating characteristic curve (AUC), accuracy, precision, sensitivity (recall), specificity, and F1-score. Feature importance was extracted from tree-based models (RF and XGBoost) to quantify the relative contribution of each predictor and enhance model interpretability.

3. Results

3.1. General characteristics

This study included 328 hospitalized patients who underwent internal fixation for DRF. Among them, 71 patients were diagnosed with poor wrist function recovery at 6 months postoperatively, yielding an incidence rate of 21.56%. Following stratified random sampling principles, subjects were grouped in a 7:3 ratio. Specifically, 230 patients were assigned to the modeling group for predictive model construction and preliminary analysis; the remaining 98 patients were allocated to the validation group to independently assess the model’s generalization capability and stability.

3.2. Independent risk factors for poor wrist function recovery in patients after internal fixation of distal radius fractures

In the modeling group, 12 variables were evaluated via univariate logistic regression analysis. Results indicated that 6 variables were closely associated with poor wrist function recovery after internal fixation of DRF: age, diabetes, postoperative pain score at 1 week, albumin level at admission, patient compliance, and fracture type (see Table 1). Further multivariable logistic regression analysis confirmed 5 independent risk factors: low compliance (OR, 2.487; 95% CI, 1.754–3.982), age > 65 years (OR, 1.725; 95% CI, 1.276–2.775), fracture type C (OR, 2.237; 95% CI, 1.773–3.761), postoperative pain score >3 points (VAS score >3 points) (OR, 1.611; 95% CI, 1.236–2.051), and decreased serum albumin (OR, 1.642; 95% CI, 1.292–2.363) (see Table 2). These factors were identified as independent risk factors for poor wrist function recovery following internal fixation of DRF, providing important clinical guidance.

Table 1.

Univariate analysis of wrist function recovery after internal fixation of DRF.

Factors	Poor recovery (N = 50)	Good recovery (N = 180)	P
Gender
Male	21	66	.491
Female	29	114
Age
≤65 yr	19	106	.009
>65 yr	31	74
BMI
≤24 kg/m2	33	135	.205
>24 kg/m2	17	45
Smoking
Yes	15	42	.334
No	35	138
Regular alcohol consumption
Yes	14	38	.303
No	36	142
Hypertension
Yes	22	65	.309
No	28	115
Diabetes
Yes	17	36	.038
No	33	144
Postoperative 1-wk VAS score
≤3	19	99	.033
>3	31	81
Admission hemoglobin level
Normal	28	120	.164
Decreased	22	60
Admission albumin level
Normal	29	135	.019
Decreased	21	45
Compliance
High	19	101	.023
Low	31	79
Fracture type
A/B	32	147	.008
C	18	33

BMI = body mass index, DRF = distal radius fractures, VAS = visual analogue scale.

Table 2.

Multivariate analysis of wrist function recovery after internal fixation of DRF.

Factors	β	SE	Waldχ2	P	OR	95% CI
Low compliance	0.911	0.176	26.792	.004	2.487	1.754–3.982
Age > 65 yr	0.545	0.311	3.071	.026	1.725	1.276–2.775
Fracture type C	0.805	0.277	8.446	.001	2.237	1.773–3.761
VAS score >3 points	0.477	0.327	2.128	.036	1.611	1.236–2.051
Decreased albumin	0.496	0.301	2.715	.037	1.642	1.292–2.363

DRF = distal radius fractures, VAS = visual analogue scale.

3.3. Comparison of machine learning model predictive performance

Based on independent risk factors identified through multivariate logistic regression analysis, this study constructed 3 predictive models for poor wrist function recovery: RF, XGBoost, and backpropagation neural network (BP). Their performance was systematically evaluated on a test dataset. Results indicated that the RF model performed best, with an AUC of 0.939, accuracy, precision, sensitivity, and specificity of 0.946, 0.970, 0.954, and 0.944, respectively, and an F1-score of 0.936, demonstrating significant advantages across multiple core metrics. The BP neural network model ranked second with an AUC of 0.924, accuracy, precision, sensitivity, and specificity of 0.933, 0.955, 0.938, and 0.929 respectively, and an F1-score of 0.913. It demonstrated robust overall predictive capability and outperformed the XGBoost model in handling structured clinical variables. In contrast, the XGBoost model achieved an AUC of 0.876, indicating weaker discrimination capabilities; Nevertheless, its classification performance remains acceptable, with accuracy, precision, sensitivity, and specificity at 0.894, 0.901, 0.909, and 0.881 respectively, and an F1-score of 0.889. This indicates it retains some value for clinical applications, though its overall performance falls short of the RF and BP models (see Table 3).

Table 3.

Comparison of predictive performance among the 3 models.

Model	AUC	Accuracy	Precision	Sensitivity	Specificity	F1-score
Random forest	0.939	0.946	0.970	0.954	0.944	0.936
BP neural network	0.924	0.933	0.955	0.938	0.929	0.913
XGBoost	0.876	0.894	0.901	0.909	0.881	0.889

AUC = area under the curve, BP = backpropagation, XGBoost = extreme gradient boosting.

3.4. Feature importance analysis

To further analyze the relative contribution of each clinical feature in predicting poor wrist function recovery, this study extracted variable importance rankings based on the RF and XGBoost models. Results showed highly consistent ranking trends across both tree models, indicating low compliance as the most influential predictor with a contribution far exceeding other variables. Secondarily, C-type fractures and age over 65 were assessed as important predictive features with substantial weights. Additionally, factors such as postoperative pain scores exceeding 3 points and low serum albumin levels demonstrated notable contributions within the model. Sensitivity analysis results from the BP neural network aligned with the tree models’ trends, further validating the stability and clinical significance of these risk factors in prediction.

4. Discussion

DRF rank among the most common fracture types, particularly in the elderly population, where their incidence increases significantly with age. These fractures typically result from falls or direct external forces applied to the wrist region.^[11] As DRFs directly impact wrist joint function, poor recovery often significantly affects patients’ daily activities and quality of life. With advances in medical technology, internal fixation surgery has become a common treatment for such fractures, effectively restoring stability at the fracture site. However, despite the positive outcomes in fracture healing achieved through internal fixation, significant individual variation persists in postoperative wrist function recovery.^[12]

This study retrospectively analyzed clinical data from 328 patients who underwent internal fixation for DRF to identify independent risk factors influencing wrist function recovery. The findings revealed that poor patient compliance, age over 65 years, fracture type C, postoperative pain scores exceeding 3 on the visual analogue scale (VAS), and decreased serum albumin levels constitute independent risk factors for poor wrist function recovery.

Building upon this, the study further introduced 3 machine learning models – RF, XGBoost, and BP neural network – to predict the risk of poor wrist function recovery and compare their performance. Results indicate that all 3 models demonstrate good predictive capability, but RF and BP neural network models outperform XGBoost overall. Among them, RF achieves the highest AUC value and stands out for its comprehensive performance, suggesting that tree models possess stronger nonlinear fitting capabilities and higher generalization performance when handling multidimensional structured clinical variables. The BP model follows closely in predictive capability, while XGBoost exhibits relatively limited discrimination ability. These findings provide clinicians with an intelligent prediction method based on objective variables, aiding in the early identification of high-risk patients for poor wrist function recovery after distal radius fracture internal fixation. This offers reliable evidence for precision rehabilitation and individualized interventions.

In this study, poor patient compliance was identified as an independent risk factor for poor wrist function recovery following distal radius fracture internal fixation. Patients with poor compliance typically fail to adhere to rehabilitation training, pain management, and early weight-bearing activities as prescribed in their treatment plans. This leads to restricted joint range of motion, weakened muscle strength, and may even result in postoperative complications such as nonunion or joint stiffness.^[13] Psychological factors like anxiety and depression also significantly impact treatment adherence. Patients lacking trust in the treatment or motivation often fail to actively participate in rehabilitation exercises.^[14] Furthermore, lack of social support or financial strain hinders sustained treatment access, further diminishing compliance. Research^[15] indicates that patients with poor compliance typically experience prolonged recovery periods and significantly poorer functional outcomes. Therefore, enhancing patient compliance – particularly through improved psychological support, adequate social assistance, and individualized treatment plans – is crucial for postoperative wrist function recovery.

In this study, age >65 years was identified as an independent risk factor for poor postoperative wrist function recovery following internal fixation of DRF. With advancing age, bone density gradually declines and fracture healing capacity markedly weakens, presenting greater challenges for elderly patients during postoperative recovery.^[16] Elderly patients exhibit slower fracture healing rates and are more susceptible to complications such as nonunion and infection, leading to poor functional recovery. Concurrently, chronic conditions like diabetes and hypertension are common in this demographic, potentially further delaying fracture healing and negatively impacting postoperative rehabilitation.^[17] Additionally, elderly patients often have reduced muscle strength, limited joint mobility, and exhibit poorer compliance and efficacy in rehabilitation exercises. Research^[18] indicates that patients over 65 years old experience slower postoperative wrist function recovery and higher rates of postoperative complications. Therefore, personalized rehabilitation plans and early intervention are particularly crucial for this population to improve postoperative recovery outcomes and enhance quality of life.

In this study, fracture type C was identified as an independent risk factor for poor wrist function recovery following internal fixation of DRF. According to AO classification, type C fractures are complex, typically involving displacement, articular surface damage, and multiple bone segments. Such fractures are generally more severe than Type A or B fractures, entailing a more complex recovery process and a higher risk of postoperative complications.^[19] Type C fractures typically require more intricate surgical management and impose higher demands on postoperative internal fixation devices. Loosening of the internal fixation or displacement at the fracture site may lead to nonunion and functional impairment.^[20] Furthermore, Type C fractures frequently involve articular surface damage, potentially causing joint instability or degenerative changes that exert long-term impacts on functional recovery. The presence of articular surface damage not only increases postoperative rehabilitation difficulty but may also lead to postoperative pain and restricted range of motion, severely impairing wrist function restoration.^[21] Research^[22] indicates that patients with C-type fractures exhibit higher postoperative pain scores, poorer joint function recovery, and longer rehabilitation periods. Consequently, surgical management for C-type fractures requires greater precision, and postoperative rehabilitation protocols should be more individualized to address their complex fracture characteristics and recovery challenges.

In this study, postoperative pain scores (VAS score > 3 points) were identified as an independent risk factor for poor wrist function recovery following internal fixation of DRF. The VAS is a commonly used method for assessing pain intensity, where higher scores indicate more severe pain. A VAS score exceeding 3 points signifies that patients experienced relatively significant postoperative pain, which adversely affects the recovery process. Postoperative pain not only directly affects patient comfort and quality of life but may also lead to activity limitations, thereby impairing recovery of joint range of motion and muscle strength.^[23] Prolonged severe pain can induce psychological barriers during rehabilitation, such as anxiety and depression, further reducing patient compliance and rehabilitation outcomes. Research^[24] indicates that patients with severe pain often resist necessary rehabilitation exercises and may even avoid weight-bearing or movement. This exacerbates issues like joint stiffness and muscle atrophy, ultimately leading to poor functional recovery of the wrist. Additionally, pain may trigger other complications, such as chronic pain syndrome or arthritis, further prolonging the recovery period. High VAS scores correlate closely with postoperative functional impairment, joint instability, and the occurrence of postoperative complications. Therefore, timely and effective pain management is crucial for improving functional recovery in patients following distal radius fracture surgery. Appropriately controlling pain and alleviating postoperative discomfort can help patients better engage in rehabilitation exercises, thereby promoting functional restoration.

In this study, decreased serum albumin was identified as an independent risk factor for poor wrist functional recovery following internal fixation of DRF. Serum albumin serves as a vital indicator of nutritional status and liver function, and is a key biomarker for assessing a patient’s physical recovery capacity. Low albumin levels commonly occur in malnutrition, chronic diseases, or acute inflammatory responses, and are closely associated with impaired immune function, impaired wound healing, and increased postoperative complications. Low serum albumin levels directly impact the fracture healing process.^[25] Research indicates that albumin not only participates in protein synthesis and metabolic processes within the body but also plays a vital role in transporting amino acids and other nutrients essential for fracture healing.^[26] Low albumin levels diminish tissue repair capacity, slow fracture healing, and increase the risk of postoperative nonunion and infection. Particularly during the postoperative recovery phase, patients with low albumin levels are more prone to muscle atrophy and joint stiffness, thereby impairing wrist function recovery.^[27] Furthermore, low serum albumin correlates with postoperative inflammatory responses. Decreased albumin levels often indicate chronic or acute inflammation, where inflammatory mediators can interfere with fracture healing and tissue repair. Research^[28] indicates that patients with low albumin levels experience higher postoperative complication rates, such as wound infections and thrombosis. These complications further delay the recovery process, leading to poor functional restoration. Therefore, low serum albumin levels postoperatively represent a significant negative predictor for fracture healing. Improving nutritional status and timely albumin supplementation can promote recovery and reduce the incidence of postoperative complications.

Among the evaluated models, the RF algorithm demonstrated the highest overall discriminative ability. This may be attributable to its ensemble structure and robustness in handling nonlinear interactions within structured clinical datasets. The consistent feature importance ranking across tree-based models further supports the stability of the identified predictors. These findings suggest that tree-based machine learning approaches may provide practical advantages for early risk stratification in clinical settings.

Although this study identified key risk factors for wrist function recovery following internal fixation of DRF, certain limitations exist. First, as a retrospective study, it relies on historical clinical data, potentially subject to information gaps or recording inaccuracies that may compromise data integrity and accuracy. Second, the study was conducted at a single hospital with a relatively limited sample size, potentially restricting the generalizability of results and preventing full representation of patient populations across different regions or healthcare institutions. Furthermore, the study did not include long-term postoperative follow-up data, precluding assessment of long-term outcomes. Finally, while multivariate regression analysis identified risk factors, it did not explore interactions or underlying mechanisms among these factors. Future prospective studies could further investigate these aspects.

5. Conclusions

This study analyzed clinical data from 328 patients who underwent internal fixation for DRF, identifying 5 independent risk factors for poor wrist function recovery: poor compliance, age over 65 years, C-type fractures, postoperative pain scores exceeding 3 points, and low serum albumin levels. Based on this, RF, XGBoost, and BP neural network models were constructed. All models demonstrated good predictive capability, with RF and BP neural network models exhibiting stronger discrimination and stability, indicating the high application value of machine learning methods in predicting the risk of poor wrist function recovery. These findings not only expand the understanding of risk factors for poor wrist function recovery after internal fixation of DRF but also provide clinicians with a convenient, individualized risk assessment tool.

In summary, the occurrence of poor wrist function recovery after internal fixation of DRF is influenced by multiple factors. Early identification and intervention for high-risk individuals are crucial for improving outcomes and optimizing healthcare resource allocation. Future multi-center, large-sample studies can provide external validation to further enhance the model’s generalization ability and clinical applicability.

Author contributions

Conceptualization: Li-ping Deng.

Data curation: Li-ping Deng.

Formal analysis: Li-ping Deng.

Investigation: Juan Li.

Methodology: Juan Li.

Software: Juan Li.

Validation: Tao Ma.

Visualization: Tao Ma.

Writing – original draft: Yi Deng.

Writing – review & editing: Yi Deng.