Responsible Software Systems for Disease Diagnostics Using Symptom Text

Abstract

Responsible software systems in digital health aim to prioritize ethical considerations, user safety, data privacy, transparency, inclusivity, accountability, sustainability, and interoperability throughout their design, development, and deployment processes. In this work, we demonstrate the use of advanced Natural Language processing techniques to build responsible intelligent systems with disease diagnostic capabilities based on symptom analysis of medical text. This involves medical language inspection based on responsible data and AI practices during extraction, normalisation and classification of diseases based on unstructured symptom text data. The core computational problems addressed in this paper include computational ambiguity arising from diverse disease symptom descriptions, untraceable temporal dependencies among extracted symptom features and the computational uncertainties among blackbox classification models. These were addressed through pre-processing medical text data to capture and incorporate temporal dependency information on disease progression, rigorous analysis and derivation of local model-agnostic interpretable for software systems transparency using bidirectional encoder representations from transformers and support vector machines. The SVM proved more convenient and efficient enough to out-compete a pre-trained BERT model with an 83.9% and 79.7% accuracy respectively on 5630 and 1410 symptom entries for training and testing respectively. We then demonstrated the possibility and feasibility of building Responsible Software Systems for reliable disease diagnostics based on text-based symptom analysis by deploying the best model. Therefore, this work also provides benchmarks for regulating digital healthcare spaces driven by intelligent software systems.