Explainable Depression Detection Using Handwriting Features

Abstract

Depression is considered one of the most prevalent diseases worldwide, with a rapid increase in recent years. An interesting area of research for depression detection is the analysis of handwriting and drawing. Although machine learning models have shown promising results to support the diagnostic process in several fields, their lack of transparency inhibits their actual use. For this reason, the work aims to develop explainable machine learning models for depression detection using handwriting-based features. The research involved 138 participants, equally divided into healthy and sub-clinical groups, according to their score on a DASS-21 (Depression, Anxiety and Stress Scale). The same protocol, consisting of seven handwriting and drawing activities, was submitted to each participant. Decision Tree and XGBoost algorithms were compared with the explainable-by-design Entropy-based Logic Explained Network (e-LEN). A 10-repeated tenfold cross-validation was employed for performance evaluation. XGBoost showed a higher AUROC 0.750±0.134 compared with e-LEN 0.723±0.143 and DT 0.681±0.119. However, e-LEN enabled a significant reduction in complexity compared with XGB. In particular, the e-LEN model exploits on average 41 logic predicates to perform the predictions, while XGBoost employs 303 nodes on average. Moreover, the e-LEN model enabled an explanation by providing the logic rules for clinician model validation. Specifically, ductus, time, and pressure features were more predictive. It is therefore possible to use these techniques and methodologies to speed up and improve the identification of depression.