AI-driven Test-Free Prediction of ICU Admission, Insulin Dependence, and Exocrine Dysfunction after Acute Pancreatitis
Ishanu Chattopadhyay, Dmytro Onishchenko, Philip Kern, Darwin ConwellObjectives:
Acute pancreatitis (AP) has heterogeneous trajectories: some patients deteriorate rapidly and require ICU care, whereas others develop delayed sequelae such as exocrine pancreatic dysfunction (EPD) and pancreatogenic diabetes. Existing scoring systems are burdensome, not available at first presentation, and poorly suited for forecasting longer-term outcomes.
Methods:
Our AI platform operating exclusively on routinely collected EHR predict three outcomes after a first recorded AP diagnosis: (i) ICU admission (same-day, within 1 wk, within 2 wk), (ii) incident EPD, and (iii) incident insulin dependence among patients without prior diabetes diagnosis or anti-hyperglycemic prescriptions. Models were trained and validated using a U.S. administrative claims database comprising 164 million individuals.
Results:
We demonstrate Area Under the Receiver-operating curve (AUC) of 0.986 (same-day ICU), 0.933 (ICU within 1 wk), and 0.927 (ICU within 2 wk). For longer-term outcomes, AUCs were 0.913 (male) and 0.901 (female) for incident EPD, and 0.861 (male) and 0.884 (female) for incident insulin dependence. Established chronic pancreatitis risk factors (e.g., obesity, tobacco dependence) are recovered with large effect sizes and high significance, supporting epidemiologic plausibility. Negative associations were consistent with etiologic subtypes, competing risks, and healthcare utilization patterns.
Conclusions:
Using only existing coded longitudinal history