A machine learning approach to profiling anxiety risk among cancer patients on treatment

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Elastic net regression: MSE 65.486 ± 8.323, MAE 6.504 ± 0.357 (vs linear regression benchmark MSE 69.945 ± 8.131). Random forest classifier accuracy: 0.751 ± 0.015. Logistic regression ROC-AUC: 0.835 ± 0.005. kNN PR-AUC: 0.797 ± 0.016. Participation self-efficacy identified as most consistent key predictor across models.

• Primary Outcome: Prediction of anxiety severity (continuous) and classification of high versus low anxiety in cancer patients on treatment

• Nnt Or Sensitivity: ROC-AUC 0.835 (logistic regression) and PR-AUC 0.797 (kNN) for binary anxiety classification; sensitivity and specificity at specific thresholds not reported; no NNT calculable from available data

• Confidence Interval: Not reported as formal 95% CIs; variability expressed as standard deviations across five imputed datasets only

Clinical Application

The identified predictors (participation self-efficacy, ECOG performance status, depressive symptoms, coping self-efficacy, symptom communication barriers, psychological well-being, age, education, type of social helper) are clinically assessable using validated instruments. However, no ready-to-use clinical tool has been produced. Implementation would require prospective validation, threshold determination, and integration into electronic health record workflows. The ML models as described are not directly deployable without further development. Anxiety in cancer patients is a recognised priority in Australian oncology care. The RACGP and Cancer Australia both endorse routine psychosocial screening in cancer care settings, consistent with the National Safety and Quality Health Service Standards (NSQHS) for Comprehensive Care. Tools such as the Distress Thermometer and HADS are commonly used in Australian cancer centres. The predictors identified in this study (ECOG status, depressive symptoms, self-efficacy, social support type) align with domains assessed in Australian psychosocial screening frameworks. However, the study's US-based sample and unspecified cancer population limit direct transferability. PBS-listed anxiolytics and psychological therapies under Medicare's Better Access initiative (including CBT) would be relevant intervention pathways if a validated risk stratification tool were developed. TGA approval of any digital clinical decision support tool would be required before deployment in Australian healthcare settings under the Software as a Medical Device (SaMD) framework. Adult cancer patients currently receiving active treatment across mixed cancer types; most directly applicable to oncology outpatient and ambulatory care settings where psychosocial screening is feasible

Abstract

PURPOSE: Previous literature has identified multiple risk factors for anxiety among individuals with cancer. However, the relative importance across many interrelated variables is not clear. Further, it is unknown how combined psychosocial factors and demographic/clinical characteristics increase an individual's vulnerability to anxiety. The aims of this study are to systematically test the following: (1) the best predictors of anxiety, (2) the best-performing classifiers for distinguishing high versus low anxiety, and (3) the most informative combinations of psychosocial features and demographic/clinical characteristics that heighten an individual's vulnerability to anxiety. METHODS: Machine learning (ML) models were used. For Aim 1, we tested the predictors of anxiety using the models of regularized regressions, GAMS, and best subset selection. For Aim 2, models tested included support vector machines (SVM), random forest, nearest neighbors, and XGBoost as the classifiers to predict high versus low anxiety. For Aim 3, gradient-boosted decision tree models were trained to examine the most important combinations of candidate predictors. We used SHapley Additive exPlanations (SHAP) to improve the interpretability of our ML models. RESULTS: Across five imputed datasets, using linear regression as the primary benchmark (MSE = 69.945 ± 8.131; MAE = 6.636 ± 0.361), elastic net showed the lowest average prediction error (MSE = 65.486 ± 8.323; MAE = 6.504 ± 0.357), although differences among predictor-based models were modest. For classifiers, random forest achieved the highest mean accuracy (0.751 ± 0.015), kNN achieved the highest PR-AUC (0.797 ± 0.016), and logistic regression achieved the highest ROC-AUC (0.835 ± 0.005). Participation self-efficacy was the most consistent key predictor of anxiety across models. The most important risk combinations of demographic and psychosocial/symptom variables included age-symptom communication barriers, age-cancer coping self-efficacy, ECOG performance status-depressive symptoms, the type of helper-psychological well-being, and education-psychological well-being. CONCLUSION: Our findings may inform risk stratification in clinical settings and facilitate personalized intervention designs aiming at relieving anxiety among cancer patients.

References

1. 1.Yan, K., Shah, M., Wang, H.-L., Elliott, A., Halbert, M., & Loerzel, V. (2026). A machine learning approach to profiling anxiety risk among cancer patients on treatment. Supportive Care in Cancer. https://doi.org/10.3390/ijerph17155457