HuLP: Human-in-the-Loop for Prognosis

Predicting future outcomes is inherently more challenging than describing current conditions. In clinical settings, clinicians typically excel at diagnosing diseases using medical images and EHR but face difficulties in making accurate individual prognostic predictions about patient survivals. This is because the former is apparent, while the latter is uncertain and depends on ingesting abundant patients’ data to make accurate individual predictions. This work presents an innovative approach designed to improve the accuracy and robustness of prognostic predictions by incorporating human expert intervention into the model’s prediction process. Specifically, we develop a deep learning model that allows human expert intervention and explore whether allowing clinicians to interact with and correct the model’s predictions can enhance prognosis accuracy, particularly in situations where data is missing or incomplete. The central hypothesis is that our human-in-the-loop approach will lead to more reliable and interpretable prognostic models compared to traditional methods.