A masked autoencoder for explainable phenotypic clustering and drug–laboratory interactions detection in psychiatric electronic health records

Makatsoris, christos; Μακατσώρης, χρήστος

dc.contributor.author	Makatsoris, christos	en
dc.contributor.author	Μακατσώρης, χρήστος	el
dc.date.accessioned	2026-07-08T11:45:32Z
dc.identifier.uri	https://dspace.lib.ntua.gr/xmlui/handle/123456789/65378
dc.identifier.uri	http://dx.doi.org/10.26240/heal.ntua.33072
dc.rights	Default License
dc.subject	Ppsychiatry	en
dc.subject	Drug–laboratory test interactions	en
dc.subject	Autoencoder	en
dc.subject	Deep learning	en
dc.subject	Explainable ai	en
dc.subject	Clustering	en
dc.subject	Ψυχιατρική	el
dc.subject	Αλληλεπιδράσεις φαρμάκων και εργαστηριακών εξετάσεων	el
dc.subject	Αυτοκωδικοποιητής	el
dc.subject	Ομαδοποίηση	el
dc.subject	Εξηγήσιμη τεχνητή νοημοσύνη	el
dc.subject	Βαθιά μάθηση	el
dc.title	A masked autoencoder for explainable phenotypic clustering and drug–laboratory interactions detection in psychiatric electronic health records	en
heal.type	masterThesis
heal.classification	cdss	en
heal.dateAvailable	2027-07-07T21:00:00Z
heal.language	en
heal.access	embargo
heal.recordProvider	ntua	el
heal.publicationDate	2026-04-06
heal.abstract	Psychiatric disorders represent a major global health challenge, associated with high disability, comorbidities, and premature mortality. Pharmacological treatment in psychiatry is complicated by drug–laboratory test interactions (DLTIs), which can obscure diagnosis, delay therapy, and compromise patient safety. Existing DLTI databases and rule-based clinical decision support systems often lack longitudinal insight and fail to capture patient heterogeneity. This thesis investigates whether deep learning–based representation learning can improve DLTI detection and provide explainable patient stratification in psychiatric care. A retrospective longitudinal cohort study was conducted using electronic health records from 248 inpatients of the 1st Department of Psychiatry, Eginition University Hospital, Athens (2020–2024). Demographic, clinical, laboratory, and pharmacological data (n = 1,490 observations) were extracted and standardized (ICD-10, ATC, LOINC). A masked autoencoder was developed in PyTorch to generate latent embeddings of patient trajectories, followed by k-means clustering (k = 3). Cluster explainability was assessed using an overfitted XGBoost classifier with SHAP interpretation. Linear Mixed Effects (LME) models were then applied to examine the association between psychotropic medications and laboratory biomarkers, adjusting for confounders and intra-patient variability. The masked autoencoder achieved stable training dynamics and generated compact, well-separated clusters (Silhouette = 0.83, Davies–Bouldin = 0.09). SHAP analysis identified gender, psychiatric diagnoses (F20–F32), age, and hospitalization frequency as the primary drivers of cluster assignment. Distinct DLTI patterns emerged across subgroups. Overall, the majority of significant DLTIs were observed within antipsychotic, antidepressant, and anxiolytic classes, predominantly affecting hematological and hepatic biomarkers such as eosinophils, bilirubin, and inflammatory markers. Distinct patterns were observed across patient clusters, reflecting underlying phenotypic and pharmacological variability. Several of these findings align with known DLTIs, while others highlight potentially novel associations warranting further investigation. This study demonstrates that masked autoencoders combined with clustering and explainability methods can uncover clinically meaningful patient subtypes and reveal drug–laboratory interactions in psychiatric populations. By integrating longitudinal EHR data with interpretable deep learning, this framework provides a scalable approach for DLTI detection and personalized monitoring. Although external validation is required, the findings underscore the promise of AI-enhanced laboratory information systems in supporting psychiatric decision making and improving patient safety.	en
heal.advisorName	Nikita, Konstantina	en
heal.advisorName	Νικήτα, Κωνσταντίνα	el
heal.committeeMemberName	Nikita, Konstantina	en
heal.committeeMemberName	Stamou, Giorgos	en
heal.committeeMemberName	Voulodimos, Athanasios	en
heal.committeeMemberName	Νικήτα, Κωνσταντίνα	el
heal.committeeMemberName	Στάμου, Γεώργιος	el
heal.committeeMemberName	Βουλόδημος, Αθανάσιος	el
heal.academicPublisher	Εθνικό Μετσόβιο Πολυτεχνείο. Σχολή Ηλεκτρολόγων Μηχανικών και Μηχανικών Υπολογιστών	el
heal.academicPublisherID	ntua
heal.numberOfPages	64 σ.	el
heal.fullTextAvailability	false
heal.fullTextAvailability	false