In this thesis, LIME (Locally Interpretable Model-Agnostic Explanations) method was applied to make the decision processes of the machine learning model transparent in diagnosing Intrauterine Growth Restriction (IUGR) using fetal heart rate (FHR) and uterine contraction (TOCO) signals obtained from cardiotocography (CTG) data. Using LIME on multivariate signal data is an innovative approach in terms of increasing the interpretability of the model and providing explanations compatible with clinical interpretations in the health field. The main purpose of the study is to determine which FHR features the model focuses on to determine the risk of IUGR and to examine the compatibility of these features with clinical indicators associated with IUGR. In this direction, the importance of signal features such as variability, acceleration and deceleration are analyzed in the light of the segment-based explanations provided by LIME. Statistical tests are used to evaluate the significance of the results; thus, it is tested whether the patterns highlighted by the model in certain segments are clinically significant. This study provides a contribution to the clinical interpretability of machine learning models by expanding the applicability of the LIME method to multivariate time series data in the field of health. Combining the predictive performance of the model with its explanatory ability, this method has shown that it can produce reliable and meaningful results in the diagnosis of IUGR. The findings suggest that the model can be used as a diagnostic aid in clinical applications and can increase the confidence in artificial intelligence applications in health services.
In questa tesi, il metodo LIME (Locally Interpretable Model-Agnostic Explanations) è stato applicato per rendere trasparenti i processi decisionali del modello di apprendimento automatico nella diagnosi del Ritardo di Crescita Intrauterino (IUGR) utilizzando il ritmo cardiaco fetale (FHR) e i segnali di contrazione uterina (TOCO) ottenuti dai dati di cardiotocografia (CTG). L’uso di LIME sui dati di segnali multivariati rappresenta un approccio innovativo per aumentare l’interpretabilità del modello e fornire spiegazioni compatibili con le interpretazioni cliniche nel campo della salute. L'obiettivo principale dello studio è determinare su quali caratteristiche dell'FHR il modello si concentri per identificare il rischio di IUGR e verificare la compatibilità di queste caratteristiche con gli indicatori clinici associati all'IUGR. In questa direzione, l'importanza di caratteristiche del segnale come variabilità, accelerazione e decelerazione viene analizzata alla luce delle spiegazioni basate sui segmenti fornite da LIME. Per valutare la significatività dei risultati vengono utilizzati test statistici; in questo modo si verifica se i pattern evidenziati dal modello in determinati segmenti siano clinicamente rilevanti. Questo studio contribuisce all'interpretabilità clinica dei modelli di apprendimento automatico estendendo l'applicabilità del metodo LIME ai dati di serie temporali multivariate nel campo sanitario. Combinando la capacità predittiva del modello con la sua abilità esplicativa, questo metodo ha dimostrato di poter produrre risultati affidabili e significativi nella diagnosi di IUGR. I risultati suggeriscono che il modello può essere utilizzato come supporto diagnostico nelle applicazioni cliniche e può aumentare la fiducia nelle applicazioni di intelligenza artificiale nei servizi sanitari.
Transparent deep learning models for diagnosing intrauterine growth restriction: an explainable ai approach with fetal cardiotocography signals
Küçükkilavuz, Hatice Deniz
2023/2024
Abstract
In this thesis, LIME (Locally Interpretable Model-Agnostic Explanations) method was applied to make the decision processes of the machine learning model transparent in diagnosing Intrauterine Growth Restriction (IUGR) using fetal heart rate (FHR) and uterine contraction (TOCO) signals obtained from cardiotocography (CTG) data. Using LIME on multivariate signal data is an innovative approach in terms of increasing the interpretability of the model and providing explanations compatible with clinical interpretations in the health field. The main purpose of the study is to determine which FHR features the model focuses on to determine the risk of IUGR and to examine the compatibility of these features with clinical indicators associated with IUGR. In this direction, the importance of signal features such as variability, acceleration and deceleration are analyzed in the light of the segment-based explanations provided by LIME. Statistical tests are used to evaluate the significance of the results; thus, it is tested whether the patterns highlighted by the model in certain segments are clinically significant. This study provides a contribution to the clinical interpretability of machine learning models by expanding the applicability of the LIME method to multivariate time series data in the field of health. Combining the predictive performance of the model with its explanatory ability, this method has shown that it can produce reliable and meaningful results in the diagnosis of IUGR. The findings suggest that the model can be used as a diagnostic aid in clinical applications and can increase the confidence in artificial intelligence applications in health services.| File | Dimensione | Formato | |
|---|---|---|---|
|
thesis_hdkkilavuz_3.pdf
accessibile in internet per tutti
Descrizione: Transparent Deep Learning Models for Diagnosing Intrauterine Growth Restriction: An Explainable AI Approach with Fetal Cardiotocography Signals
Dimensione
3.11 MB
Formato
Adobe PDF
|
3.11 MB | Adobe PDF | Visualizza/Apri |
|
Executive_summary_hdkkilavuz.pdf
accessibile in internet per tutti
Descrizione: Executive Summary
Dimensione
1.06 MB
Formato
Adobe PDF
|
1.06 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/230861