Research Scientist, Medical World Models

Function Health•Canada, KS

1d•Remote

About The Position

Function Health is building a multimodal, longitudinal picture of health for hundreds of thousands (soon to be millions) of people. The Function proactive health dataset includes whole-body MRI, 100+ blood biomarkers, radiology reports, questionnaires and, increasingly, wearables data, all linked to the same individual and refreshed over time. The Medical Intelligence Lab (MIL) at Function exists to turn that data into an early-warning system for every member. The World Model team’s job is at the core of that ambition. We are building models that can do two things: represent a member’s health today from whatever data exists historically, and predict how that state evolves, what the next lab panel is likely to show, what the next scan is likely to find, and eventually how that trajectory changes under an intervention. As a Research Scientist on this team you will be a key technical contributor, working directly with the Team Lead and MIL’s Chief Medical Scientist. You will design and train the models, own the evidence that they work, and ship them as the pretrained foundation upon which a portfolio of MIL products and tools build. The work is expected to reach large numbers of members and also the scientific literature; we publish, and we deliver.

Requirements

PhD in machine learning, computer science, biomedical engineering or a related field with 1-2+ years of professional experience, or MS/BS with 5+ years building and evaluating ML models on real data.
Demonstrated depth in at least one of: temporal / longitudinal modelling (sequence models, neural ODE/CDE or state-space models, forecasting with irregular sampling, survival or progression modelling); self-supervised or foundation-model pretraining on medical imaging (3D MRI/CT) or multimodal data; multimodal fusion of imaging with tabular, EHR or biomarker data.
Strong Python and PyTorch; experience training at scale (multi-GPU, large datasets, experiment tracking) and writing code others build on.
Rigorous evaluation instincts: statistical thinking, calibration, error analysis, and the habit of asking whether a result would survive a larger validation set.
Publication record at top ML or medical-imaging venues (e.g., MICCAI, NeurIPS, ICML, ICLR, CVPR), or equivalent evidence of research output delivered into production.
Clear written and verbal communication with technical and clinical audiences.

Nice To Haves

Experience with world models, latent dynamics, model-based RL, or counterfactual / causal inference on observational health data.
Work with longitudinal cohorts or biobank-scale data (e.g., UK Biobank, NAKO, ADNI) or with EHR/lab time series.
Tabular foundation models or numeric tokenization for continuous clinical values; normative modelling; biological/organ-age estimation.
Vision–language pretraining with radiology reports; report information extraction with LLMs.
Cloud ML infrastructure (AWS, Databricks); experience in healthcare or other regulated, PHI-sensitive environments.
Prior experience as a founding or early member of a research team.

Responsibilities

Design and train models that forecast a member’s next health state from irregularly sampled histories (repeat biomarker panels, questionnaires, and imaging-derived features, etc) with calibrated uncertainty at clinically meaningful horizons.
Develop encoders capable of unifying medical data into a fused representation that is robust to arbitrary missing modalities, and that transfers efficiently to downstream clinical tasks.
Build and maintain the evaluation framework that decides what we ship: powered validation sets, confidence intervals, label-efficiency curves, forecasting metrics, collapse diagnostics, and benchmark registration against MIL’s clinically validated specialist models.
Own the training and evaluation codebase end to end: data loaders over our standardized research exports, distributed training on GPU infrastructure, experiment tracking, versioned model releases with model cards and licence inventories.
Deliver pretrained encoders and forecasters to MIL’s product-track teams as documented, reproducible artefacts, and support their integration through clinical validation and hand-off to Engineering and Product Development.
Work with the Data Team on dataset specifications, and requirements.
Differentiate between predictive and causal claims: we forecast from observational data and validate before we assert. Design analyses so that limitations are explicit and reviewable.
Work within a PHI-sensitive, regulated environment: de-identified data only, licensable pretrained weights/data, documentation that meets regulatory review.
Write it down: experiment logs, design notes, and the periodic “what we learned” retrospectives that shape the roadmap. Publish at venues such as NeurIPS, ICML, ICLR, CVPR, ICCV/ECCV, AAAI, MICCAI etc in order to establish the lab.
Collaborate daily with ML engineers on adjacent MIL projects, the MIL Data, Infrastructure, and Quality and Regulatory Affairs teams, and clinicians on the Medical Integration Team.
Help define how this team works: research reviews, documentation standards, code review, and help hire the next members.