Ali Namvar, Assistant Research Scientist, Radiology, Michigan Medical School
Watch RecordingAli Namvar introduced himself as “a molecular biologist in a room full of engineers,” and that displacement turned out to be generative rather than paralyzing. The mismatch in backgrounds forced him to articulate, in precise terms, what he actually needed from AI, rather than reaching for whatever method was most familiar.
The problem he inherited in the ICU was not a classification or prediction problem in the conventional sense. It was a whole-picture problem. In a high-acuity setting, data streams in continuously from everywhere (bedside monitors, lab results, physician notes), and existing tools either compress all of it into a single score or fire alarms without conveying overall direction. What was missing was a system that could look at the features together, respect the relationships between them, and track how a patient’s physiological state was moving over time. “We follow the question, not the tools,” Namvar said. That reorientation led him to optimal transport, a mathematical framework that measures the minimum “work” required to reshape one probability distribution into another. It does not compare averages. It captures movement. “Optimal transport was a natural fit.”
STREAM’s most consequential design choice was to let the clinical states emerge from the data rather than defining them in advance. Trained on the EICU dataset of nearly 158,000 patients, the system identified five distinct physiological states, each with its own mortality profile and signature, without any labels. But the finding that most surprised the team was not the states themselves. It was the outliers: patients assigned to a state who did not fit it well. Time spent as a “state outlier” was associated with nine times higher mortality risk than for patients who consistently fit their assigned state. This was not something the team set out to find. It emerged from the analysis, and they had to validate it on an entirely independent dataset before trusting it.
That validation came from MIMIC, collected from a different hospital system entirely, and it reproduced the effect: state outliers carried seven times higher mortality risk, and the mortality gradients across the five states held between datasets. The EICU model was frozen and applied to MIMIC without retraining, and the structure it had learned held. External validation, Namvar argued, is not a final step tacked onto a finished project. It is a design requirement: the difference between a system that detected something about one dataset and one that detected something about patient physiology.
The broader lesson was about team composition as much as methodology. “It’s not a one-person or one-discipline project.” Clinical judgment was needed to decide which features mattered, engineering to build a pipeline stable enough for production data, and a biology mindset, treating the patient as a system rather than a set of isolated measurements, to frame the problem correctly in the first place. Data quality, he emphasized, is where all of it stands or falls. “If you skip the pre-processing, no model is saving you.”