Explainable Profiling of Sleep Disorders to Support Trustworthy Clinical Interventions

Sifeddine Sellami, Juba Agoun, Lamia Yessad, Auday Berro, Louenas Bounia

Abstract:

Sleep disorders present significant diagnostic challenges due to symptom heterogeneity, profoundly impacting
public health. To address this, we leverage real-world data from
the KANOPEE mobile application—a validated digital tool for
managing sleep complaints—to develop a trustworthy model
for disorder detection and patient profiling using explainable
artificial intelligence (XAI). Guided by clinical expertise, our
analysis focuses on key variables including demographic characteristics, clinical assessment scores (e.g., Insomnia Severity Index,
anxiety, and depression scores), and physiological metrics. Our
dataset comprises approximately 1k complete patient records.
Our dual methodological approach involves: (1) regression models
to predict treatment success metrics, enabling the identification
of modifiable behavioral factors influencing mental health and
sleep outcomes; and (2) k-means clustering to segment patients
into three or more distinct profiles, differentiating subtypes of
sleepers (e.g., young professionals with mild insomnia and anxiety,
older individuals with regular sleep and low depression, and
middle-aged patients with high daytime sleepiness and irregular
sleep patterns). XAI techniques, including SHAP and formal
explanations, elucidate key drivers like sleep regularity and
anxiety, validated against clinical thresholds and expert feedback from sleep specialists. This framework aligns algorithmic
insights with medical reasoning, enhancing interpretability and
supporting personalized interventions in digital sleep medicine.
Future extensions will incorporate application completion rates
for broader efficacy assessment.