University of Twente Student Theses

Login

Explainable Diagnoses Prediction For General Practitioners

Zeilstra, Pieter (2023) Explainable Diagnoses Prediction For General Practitioners.

[img] PDF
20MB
Abstract:In this study, a Dutch variant of the RoBERTa language model known as RobBERT was made domain-specific to the domain of Dutch general practitioners(GPs). The model was trained and fine-tuned using 2.2 million user-identified symptoms (S-rules) derived from SOAP(Subjective, Objective, Assessment and Plan) notes. ICPC codes were used to classify a diagnosis based on the S-rule. A new classification head was introduced, enabling the separate classification of ICPC symptom codes and ICPC disease/diagnosis codes. To align with the diagnostic decision-making process of GPs, a threshold function was implemented to determine the number of ICPC codes returned. The threshold function outperformed the simple approach of returning only the top three codes while providing fewer than three codes on average. When predicting ICPC symptom codes, the threshold function achieved an accuracy of 90%; for ICPC diagnosis/disease codes, the accuracy was 88.6%. A LIME symptom-module was proposed. The symptom-module is an adaptation of the LIME text-module. The module aimed to identify the most important symptoms for each ICPC diagnosis/disease code. Furthermore, a user study was conducted with five participating general practitioners, of which four participants found that the model contributed to their diagnostic ability by suggesting ICPC codes.
Item Type:Essay (Master)
Clients:
Topicus, Deventer, Netherlands
DigiDok, Utrecht, Netherlands
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:https://purl.utwente.nl/essays/96151
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page