Uncertainty Evaluation and Patient-Based Calibration for Early Sepsis Prediction in Contrast to Standard Machine Learning Models

Lahari Kolukuluri, Archana Kumm, Sana Sachdeva

Keywords: sepsis prediction, machine learning, uncertainty quantification, calibration, deep ensembles, Monte Carlo dropout

Abstract: Sepsis is a major worldwide cause of death, and early identification is critical in reversing patient outcomes for the better. Although machine learning models have been hopeful for sepsis prediction, these models generally sacrifice confidence estimation and calibration for accuracy, rendering them clinically less suitable. This paper proposes a new framework for early sepsis prediction with patient-specific calibration and uncertainty estimation to allow for greater trustworthiness and interpretability. We used recurrent neural network architectures with Monte Carlo dropout, deep ensembles, and probabilistic output heads to the PhysioNet Sepsis Challenge dataset of time-series vital signs, lab tests, and demographics. Calibration techniques such as temperature scaling and conformal prediction were used to map predicted probabilities into observed outcomes across subgroups of patients. Experimental results validate that the proposed systems possess discriminative superiority and lower ECE considerably, along with yielding informative confidence estimates to recognize low-trust predictions. By successfully combining accurate prediction and reliable uncertainty estimation, this paper solves constructing clinically acceptable machine learning systems for diagnosing sepsis in its early stage.

Submission Number: 49