Bayesian Networks in the Management of Hospital Admissions: A Comparison between Explainable AI and Black Box AI during the Pandemic.

Artificial Intelligence (AI) and Machine Learning (ML) approaches that could learn from large data sources have been identified as useful tools to support clinicians in their decisional process; AI and ML implementations have had a rapid acceleration during the recent COVID-19 pandemic. However, many ML classifiers are "black box" to the final user, since their underlying reasoning process is often obscure. Additionally, the performance of such models suffers from poor generalization ability in the presence of dataset shifts. Here, we present a comparison between an explainable-by-design ("white box") model (Bayesian Network (BN)) versus a black box model (Random Forest), both studied with the aim of supporting clinicians of Policlinico San Matteo University Hospital in Pavia (Italy) during the triage of COVID-19 patients. Our aim is to evaluate whether the BN predictive performances are comparable with those of a widely used but less explainable ML model such as Random Forest and to test the generalization ability of the ML models across different waves of the pandemic.

Performance of an AI algorithm during the different phases of the COVID pandemics: what can we learn from the AI and vice versa.

Background: Artificial intelligence (AI) has proved to be of great value in diagnosing and managing Sars-Cov-2 infection. ALFABETO (ALL-FAster-BEtter-TOgether) is a tool created to support healthcare professionals in the triage, mainly in optimizing hospital admissions. Methods: The AI was trained during the pandemic's "first wave" (February-April 2020). Our aim was to assess the performance during the "third wave" of the pandemics (February-April 2021) and evaluate its evolution. The neural network proposed behavior (hospitalization vs home care) was compared with what was actually done. If there were discrepancies between ALFABETO's predictions and clinicians' decisions, the disease's progression was monitored. Clinical course was defined as "favorable/mild" if patients could be managed at home or in spoke centers and "unfavorable/severe" if patients need to be managed in a hub center. Results: ALFABETO showed accuracy of 76%, AUROC of 83%; specificity was 78% and recall 74%. ALFABETO also showed high precision (88%). 81 hospitalized patients were incorrectly predicted to be in "home care" class. Among those "home-cared" by the AI and "hospitalized" by the clinicians, 3 out of 4 misclassified patients (76.5%) showed a favorable/mild clinical course. ALFABETO's performance matched the reports in literature. Conclusions: The discrepancies mostly occurred when the AI predicted patients could stay at home but clinicians hospitalized them; these cases could be handled in spoke centers rather than hubs, and the discrepancies may aid clinicians in patient selection. The interaction between AI and human experience has the potential to improve both AI performance and our comprehension of pandemic management.