03 / Doctoral research · King’s College London
Probabilistic simulation
of ICU trajectories
Realistic intensive-care EHR trajectories are difficult to simulate because physiology, treatments, observation times and patient-existence state co-evolve under sparse clinical measurement.
Saved PULSE CNF checkpoint outputs from synthetic, non-patient histories. Enable JavaScript to explore a model trained on CVSim-style data in three saved scenarios.
A separate PULSE/CNF model trained on synthetic CVSim-style heart rate and blood pressure. One saved simulator path per scenario, with paired treatment starting at hour 8. These experiments do not establish clinical causal effects.
Probabilistic simulation
PULSE separates treatment rollout from patient-state rollout, predicts binary patient state before continuous physiology, and updates continuous variables through gated residual dynamics around the last observed or simulated state. Probabilistic variants use heteroskedastic Gaussian, monotone B-spline or continuous-flow residual distributions.
Baseline and evaluation
The work includes reconstructing a published transformer baseline, developing Gaussian and normalising-flow output heads, and examining calibration and trajectory behaviour. Patient-level split discipline and reproducible evaluation accompany model development.
MIMICWizard
I’ve extended a fork of MIMICWizard with patient-timeline viewing, ICU filtering and cohort-related SQL support.
Explore my MIMICWizard forkThese saved forecasts use synthetic inputs. The cardiovascular generator is a mechanistic-inspired CVSim-style benchmark; clinical forecasting, treatment efficacy and original-CVSim physiological validation remain outside this demonstration.
Up next / Applied AI