Machine-Learning-based Prioritisation of Long COVID Patients for Specialist Consultation

Abstract

Introduction: Long COVID, a post-acute sequela of SARS-CoV-2 infection, poses a significant burden on healthcare systems. Efficient triaging of patients for specialist consultation is vital for timely and appropriate care. A machine learning (ML) algorithm is proposed in this thesis to prioritise long-COVID patients for initial consultations with cardiologists, pulmonologists, or psychiatrists. Methods: A ML model was developed using a Support Vector Machine (SVM) classifier with Synthetic Minority Oversampling Technique (SMOTE) for class imbalance handling and feature selection techniques. The dataset was acquired from an ongoing research project that examines holistically Long Covid patients. Data, collected from the first consultation in the Long Covid clinic, from 175 patients included demographics, COVID-19 severity, vaccination status, validated questionnaires (BECK Depression, PTSD, ABC, CAT, EQ-5D, HADS Depression/Anxiety, SF-36), and specific symptoms (intestinal disorders, sleep disorders, precardial pain). The model was trained and validated using 5-fold Stratified Cross-Validation and shuffling of the data. Results: The SVM model achieved an accuracy of 0.67 and an Area Under the Curve (AUC) of 0.84, outperforming other models (Random Forest, K-Nearest Neighbours, Decision Tree, XGBoost). Notably, the SF-36 General Health (SF-36 GH-N) score emerged as the most important factor for patient risk group classification, while SF- 36_BP-N (bodily pain), and the HADS anxiety questionnaire follow. Conclusion: This study demonstrates the potential of an ML-based approach to prioritise long-COVID patients for specialist consultations. By leveraging patient data and validated questionnaires, the algorithm can guide resource allocation and expedite access to appropriate specialists, improving patient care. The results suggests that a patient's overall health perception plays a crucial role in identifying the most needed specialist for their initial consultation. Future work will investigate the impact of the algorithm on clinical workflow and patient outcomes in a prospective study.