Talking with the Doctor: Understanding and Communicating Models Performance in Healthcare

Effectively communicating the results of predictive models to doctors remains challenging. This study focuses on improving communication between data scientists and medical staff by analyzing the classification of COVID-19 patients into severe and non-severe groups. By measuring instance hardness and examining meta-features, we identified key patterns for classification. Severe patients were easily classified when aged 64+ and presented low lymphocyte percentage (LYM%), while non-severe patients exhibited high LYM% or low LYM% with low urea values. Meta-features provided additional insights into model decision-making. These findings enhance communication, enabling doctors to better interpret model’s outputs. By simplifying features and meta-feature presentation and emphasizing interpretability, collaboration between data scientists and doctors is facilitated, opening a pathway to build data-centered models.