An interpretable machine learning framework for screening psychological distress in patients with chronic
Haomiao Zhao1, Xuanchen Zhou2, Xiaojun Zhang1
1Department of Otorhinolaryngology, National Health Commission Key Laboratory of Otorhinolaryngology, Shandong Provincial Key Medical and Health Discipline, Qilu Hospital of Shandong University, Jinan, Shandong 250061, China.
Background:
Chronic rhinosinusitis (CRS) is associated with substantial psychological burden, yet tools for identifying patients with psychological distress are lacking in otolaryngologic practice.
Objective:
To develop and validate an interpretable machine learning model for screening psychological distress in patients with CRS.
Methods:
This multicenter study used preoperative data from 408 adults with CRS undergoing endoscopic sinus surgery and an independent external validation cohort of 61 patients. Psychological distress was assessed using the Hospital Anxiety and Depression Scale. After feature selection, nine machine learning models were compared in training and internal test sets. The final model was applied to the external cohort without refitting. Performance was evaluated using discrimination, calibration, and decision curve analysis (DCA), with interpretability assessed using SHapley Additive exPlanations.
Results:
Psychological distress was present in 42.9% of the development cohort and 37.7% of the external cohort. Key predictors were SNOT-22 score, nasal congestion, nasal polyps, antihypertensive drug use, and ear fullness. CatBoost showed the best performance, achieving an AUC of 0.864 (95% CI, 0.793-0.927) in the internal test set, with a Brier score of 0.146. DCA showed the model outperformed the treat-all and treat-none strategies across a broad range of threshold probabilities. In external validation, the model achieved an AUC of 0.792 (95% CI, 0.653-0.913). The final model was translated into a web-based tool for individualized risk estimation.
Conclusion:
Psychological distress is common in CRS. An interpretable CatBoost model using routine clinical variables showed good discrimination and external validity, supporting its potential use for screening and risk stratification.