DataScience SG x ODSC Meetup - Applying ML to Healthcare

Eugene Yan spent nearly a year and a half inside a healthtech startup called uCare.ai before finally giving a talk about it at a DataScience SG x ODSC meetup in Singapore. The project itself is the kind of thing that sounds mundane until you realize how much it affects real people: a system that predicts a patient's total hospital bill before they're even admitted, deployed inside Parkway Pantai, the largest hospital group in Southeast Asia.

The motivation isn't just convenience. Singapore's Ministry of Health actually requires this kind of pre-admission cost transparency, and for good reason — nobody wants a surprise bill at discharge. Yan's team replaced Parkway's existing estimation system with a machine learning pipeline that cut mean absolute error by 55% and root mean squared error by 60%. Just as important, they tuned the model to lean toward overestimating rather than underestimating, since a slightly high quote manages expectations far better than a lowball number that turns into a shock later.

What struck me most about Yan's account is the rollout itself. The front-end interface stayed exactly the same for hospital staff — the backend just started calling a new API instead of the old estimation logic. No retraining staff, no new software to learn. Hospital administrators reported virtually zero complaints after the switch, which is the kind of quiet success story that never makes headlines but actually matters to patients standing at a billing counter.

Yan's bigger point, though, is about where the effort in a project like this actually goes. Data engineers handled encryption and ETL pipelines. DevOps packaged and deployed the models. Infrastructure teams set up cloud storage, compute, and networking. The actual machine learning — the part everyone assumes is 80% of the job — made up roughly 20% of the total work. He argues this inverts what most newcomers to data science, and plenty of academics, assume about the field.

He also takes a swipe at Kaggle-style thinking, where competitors chase marginal accuracy gains on clean CSVs using complex ensembles that take days to train. Useful for exploring ideas, he says, but largely disconnected from what production machine learning actually demands: proper framing of the problem and solid engineering for deployment and long-term maintenance. That's the gap he's hoping more Southeast Asian companies start to close.