Technical Summary: Development and Validation of a Machine Learning-Based Model for Predicting Frailty Risk in Head and Neck Cancer Patients Undergoing Radiotherapy

Problem Statement

Frailty syndrome, characterized by reduced physiological resilience and increased susceptibility to stressors, represents a critical clinical challenge for patients with head and neck cancer (HNC) undergoing radiotherapy. While radiotherapy improves survival, its adverse effects (e.g., mucositis, dysphagia) can exacerbate frailty, leading to reduced treatment tolerance, poorer quality of life, and increased mortality. Current assessment methods rely on clinician judgment and static scales, which are often time-consuming, subject to inter-rater variability, and lack the ability to holistically integrate multidimensional data (clinical, psychological, and social). There is a distinct lack of effective, predictive tools capable of early identification of at-risk patients to facilitate targeted interventions.

Methodology

Study Design and Participants

This cross-sectional study enrolled 698 HNC patients undergoing radiotherapy at a tertiary grade A hospital in Sichuan Province, China, between February 22, 2025, and July 14, 2025. Participants were selected via convenience sampling. Inclusion criteria required confirmed HNC diagnosis, active radiotherapy, clear consciousness, and signed informed consent. Exclusion criteria included unstable vital signs, cognitive/psychiatric disorders, or major organ dysfunction. The sample size (n=698) was calculated based on the Events Per Variable (EPV) principle, exceeding the minimum requirement of 472 cases derived from a 42.4% expected frailty prevalence.

Data Collection and Instruments

Data were collected using five primary instruments:

1. General Information Questionnaire: Captured sociodemographics, clinical characteristics (cancer stage, pain, nutritional risk), and laboratory parameters (hematology, albumin, creatinine, etc.).
2. Charlson Comorbidity Index (CCI): Assessed comorbidity burden.
3. Pittsburgh Sleep Quality Index (PSQI): Evaluated sleep quality.
4. Hospital Anxiety and Depression Scale (HADS): Measured anxiety and depression levels.
5. Perceived Social Support Scale (PSSS): Assessed social support domains.
6. FRAIL Scale: Used as the outcome measure to classify patients as robust, pre-frail, or frail. For analysis, robust and pre-frail categories were consolidated into a "non-frail" group.

Data Preprocessing and Feature Selection

Data were processed in a Python 3.9 environment (Jupyter Notebook 7.2.2). Missing values were handled via multiple imputation (modal substitution for categorical, mean imputation for continuous). Outliers were detected using Isolation Forest and replaced with feature medians. Features were normalized (min-max scaling) or label-encoded.
To mitigate overfitting and select optimal predictors, LASSO (Least Absolute Shrinkage and Selection Operator) regression was employed with 10-fold cross-validation. This process reduced the initial candidate pool to 17 clinically significant predictors, including age, occupation, income, sleep disorders, anxiety, depression, clinical stage, pain, nutritional risk, and specific laboratory indices (ALB, PLT, TG, hypokalemia).

Model Development and Validation

The dataset was stratified into a training set (70%, n=488) and a validation set (30%, n=210). Ten machine learning algorithms were constructed and compared:

• Logistic Regression (LR)
• Support Vector Machine (SVM)
• Random Forest (RF)
• Decision Tree (DT)
• K-Nearest Neighbors (KNN)
• Extra Trees (ET)
• LightGBM
• Stochastic Gradient Descent (SGD)
• Naïve Bayes (NB)
• XGBoost

Model hyperparameters were optimized using grid search with 10-fold cross-validation on the training set. Performance was evaluated on the independent validation set using Accuracy, Precision, Recall, Sensitivity, Specificity, F1-score, and the Area Under the Receiver Operating Characteristic Curve (AUC). The optimal model underwent interpretability analysis using the Shapley Additive explanation (SHAP) method.

Key Results

Frailty Prevalence

The study identified a frailty incidence of 42.4% (296 out of 698 patients) within the cohort.

Model Performance

Among the ten algorithms tested, the Random Forest (RF) model demonstrated superior performance on the validation set, achieving:

• AUC: 0.940 (95% CI: 0.910–0.966)
• Accuracy: 0.843
• Precision: 0.889
• Recall/Sensitivity: 0.719
• Specificity: 0.934
• F1-score: 0.795

While XGBoost and LightGBM showed comparable AUCs in the training phase, the RF model demonstrated enhanced calibration precision in the validation set.

Feature Importance and Interpretability

SHAP analysis identified the most influential predictors of frailty risk. The top three predictors were:

1. Depression (HADS-D score): The most significant predictor.
2. Sleep Disorders (PSQI): A critical secondary factor.
3. Pain Severity: A key contributor.

Other significant factors included platelet count (PLT), age, triglycerides (TG), clinical stage, radiation-induced injury, anxiety, occupation, income, catheter type, disease site, comorbidity, hypokalemia, and nutritional risk. Notably, albumin (ALB) exhibited a protective negative association with frailty risk. The SHAP summary plot confirmed positive associations between frailty risk and depression, sleep disorders, pain, age, and clinical stage, while highlighting the protective role of albumin.

Significance and Claims

The authors assert that this study provides a clinically actionable, explainable tool for the early identification of HNC patients at risk of frailty during radiotherapy. The primary significance of the work lies in three areas:

1. High Discriminative Power: The developed RF model demonstrates outstanding predictive capability (AUC = 0.940), surpassing traditional assessment limitations.
2. Clinical Transparency: By integrating SHAP analysis, the study moves beyond "black box" predictions, offering transparent visualization of how specific features drive risk. This addresses the interpretability constraints often associated with complex ML models.
3. Targeted Intervention Targets: The model explicitly identifies modifiable predictors—specifically depression, sleep quality, and pain. The authors claim these findings point to key targets for clinical management, suggesting that interventions such as cognitive behavioral therapy, sleep hygiene optimization, and multimodal pain management could potentially delay or reverse frailty progression.

The study concludes that this algorithmic framework supports proactive prevention strategies, enabling clinicians to stratify risk and implement precision interventions to improve treatment success and health-related quality of life for this high-risk population. The authors acknowledge limitations, including the single-institution design and the cross-sectional nature of the study, noting that future work should focus on multi-center validation and longitudinal modeling.