Towards early psychosis and depression prediction - optimizing an explainable machine learning pipeline based on multimodal mri data

Aug 4, 2026

Upload by Skyler

Distinguishing recent-onset psychosis (ROP) from depression (ROD) remains difficult in routine psychiatric practice, where clinician judgment based on reported symptoms can lead to misdiagnosis and delayed intervention. This thesis builds an explainable, optimized machine learning pipeline using structural and functional MRI from 506 participants in the multi-site European PRONIA study, comparing healthy controls with patients and classifying ROP versus ROD. Nested cross-validation integrates confounder correction and hyperparameter optimization, while feature selection and optimization explore filter methods and genetic algorithms. Shapley values and permutation importance support interpretability. Incorporating substance-use information improves classification accuracy.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages118

Reading Time4 h 57 min

AccessFree to download

Introduction and Aim

Background

State Of the Art

Confounding variables

Biomarkers discovery and predictive models

Machine learning in diagnostic applications

Classification with machine learning techniques

Validation

Model interpretation

Towards early psychosis and depression prediction: optimizing an explainable machine learning pipeline based on multimodal mri data Tesi di Laurea Magistrale in Biomedical Engineering-Ingegneria Biomedica Author: Gianluca De Franceschi Student ID: 993890 Advisor: Prof. Eleonora Maggioni Co-advisors: Inês Won Sampaio, Prof. Paolo Brambilla Academic Year: 2023-24 i Abstract Distinguishing between recent-onset psychosis (ROP) and depression (ROD) remains a significant challenge in psychiatry. The current clinical practice relies on subjective clinician evaluation, leading to potential misdiagnosis and delayed treatment. This study ventures beyond symptom-based diagnosis by harnessing the power of neuroimaging and machine learning (ML) . We present an optimized framework to develop a ML pipeline designed to distinguish healthy controls (HC) from patients (P) and to classify ROP and ROD using structural and functional Magnetic Resonance Imaging (MRI) data from 506 participants of the multi-site European PRONIA study. Recognizing the potential influence of substance use, we explore the possibility of incorporating these additional features. To ensure model robustness and generalizability, a nested cross-validation (CV) scheme is implemented, featuring a unique design that integrates confounder correction and optimization tools. The pipeline development involved a meticulous search for the best-performing featureselection (FS) and optimization techniques including filter-based methods and genetic algorithms (GA) . Finally, Shapley values and permutation importance techniques were used to gain valuable insights into the model’s decision-making process. The GA consistently demonstrated its potential for applications in psychiatric diagnosis. Used both for feature selection and model selection tasks, it achieved the following average accuracies: 0 .574 for HC vs. P and 0 .676 for ROP vs. ROD. Our findings also revealed that incorporating data on substance use improved the overall classification accuracy. By leveraging the power of neuroimaging and ML, this study paves the way for more objective data-driven diagnoses of ROP and ROD, ultimately contributing to improved patient outcomes. Keywords: psychosis, depression, MRI, PRONIA, machine learning, genetic algorithms, shapley values, nested cross-validation Abstract in lingua italiana Il riconoscimento di psicosi o depressione ad esordio recente (ROP o ROD) rimane unostacolo significativo in psichiatria. I metodi tradizionali si basano fortemente sui sintomiriportati dai pazienti, portando a potenziali diagnosi errate e ritardi nel trattamento. Questo studio si spinge oltre la diagnosi basata sui sintomi sfruttando i concetti della diagnostica per immagini e del machine learning (ML) . Viene qui presentato un framework per lo sviluppo di una pipeline di ML ottimizzata, progettata per distinguere i pazienti (P) dai controlli sani (HC) e i casi ROP dai ROD , allenata su dati di risonanza magnetica strutturale e funzionale (RM) raccolti dal consorzio PRONIA in uno studio europeo multicentrico che comprende 506 soggetti. Riconoscendola potenziale influenza dell’abuso di sostanze, è stata esplorata anche la possibilità di incorporare il loro consumo come dato aggiuntivo. Per garantire la robustezza e la generalizzabilità del modello, è stato implementato uno schema di nested cross validation, caratterizzato da un design unico capace di integrare gli strumenti di correzione dei fattoriconfondenti ai sistemi di ottimizzazione. Una meticolosa selezione delle tecniche di feature selection e di ottimizzazione più performanti, che ha incluso filter methods e algoritmi genetici (AG), è stata condotta. Infine, i valori di Shapley e le tecniche di permutation importance sono state utilizzate per ottenere preziose informazioni sul processo decisionale del modello. Gli AG hanno dimostrato il loro potenziale per le applicazioni nella diagnosi psichiatrica; utilizzati sia

FAQs

What problem does the study address in psychiatric diagnosis?

It targets the difficulty of distinguishing recent-onset psychosis (ROP) from depression (ROD), which is often limited by subjective symptom-based clinical evaluation.

Which data and participants are used to train and evaluate the pipeline?

The pipeline uses structural and functional MRI data from 506 participants from the multi-site European PRONIA study, including healthy controls and patients.

How does the method ensure robustness and generalizability?

It uses a nested cross-validation scheme that includes a design for confounder correction and optimization tools.

Aug 4, 2026

Upload by Skyler