The UTC Graduate School is pleased to announce that Andrew Stuckey will present Master’s research titled, Early Warning Signals for Cholera Resurgence in Haiti: Evaluating Deep Learning Under Epidemic Regime Shift on 10/28/2026 at 3:30 PM in zoom. Everyone is invited to attend.
Mathematics
Chair: Xiunan Wang
Co-Chair:
Abstract:
Deep learning classifiers trained on stochastic epidemic simulations show promise as early warning signals for epidemic transitions, but whether they generalize from historical outbreak dynamics to post-elimination resurgence remains untested. We trained a CNN-LSTM classifier (convolutional neural network–long short term memory network) on simulations from three cholera models parameterized to Haiti’s 2010-2019 epidemic, then evaluated it on an independent model of Haiti’s 2022 resurgence. The cholera-trained classifier performed well in distribution but near or below chance under regime shift. A generic epidemic-model trained classifier transferred at short input windows but collapsed at longer ones, suggesting disease-specific training did not improve generalization. Training the same architecture on resurgence simulations produced near perfect discrimination, confirming the failure reflects distribution mismatch, not architecture. Lag-1 autocorrelation remained robust under regime shift, while rolling variance did not. These results suggest lag-1 autocorrelation is as reliable as the best classifier for anticipating post-elimination cholera resurgence.