The reported gains are real but measured: AUROC improvements of 0.02–0.08 over gradient boosting across mortality, readmission and diagnosis-onset tasks. On a large population that is worth having; it is not a step change.
Next-event models report top-10 accuracy around 0.68–0.78 depending on site, and that site dependence is the point — these models are more sensitive to population and coding practice than any other group on this site.
Every model on this page runs as part of a managed AI pipeline in our GPU clusters, with a dedicated private cluster in our cloud or an on-premise installation where medical governance requires it. Output is decision support for a qualified professional to review, not a diagnosis.
Models in this group take longitudinal coded record data rather than free text; the rate is patient timelines per hour. Each table gives three hardware tiers — Minimum, the smallest configuration on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is what one server of that tier processes per hour. Use these figures for initial sizing only. Before production we benchmark your own data to confirm accuracy, latency, throughput and cost.
CLMBR / MOTOR
Vendor: Stanford (Shah lab)
What it does: turns a coded patient timeline into embeddings and time-to-event risk scores, beating gradient boosting by 0.02–0.08 AUROC across mortality, readmission and diagnosis-onset — and the same embedding serves several endpoints.
AUROC gains of 0.02–0.08 over gradient boosting across mortality, readmission and diagnosis-onset tasks (independent, EHRSHOT).
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
12 GB
24 GB
80 GB
vCPUs
6
12
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (documents/hour)
≈ 6,000
≈ 21,000
≈ 54,000
Foresight / ETHOS
Vendor: Community, including Google Health-derived open reimplementations
What it does: forecasts the next clinical events with probabilities, at top-10 accuracy around 0.68–0.78. Site dependence is high, which is why local training is part of the work rather than an optional extra.
Next-event prediction top-10 accuracy about 0.68–0.78 depending on site (independent).
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
12 GB
24 GB
80 GB
vCPUs
6
12
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (documents/hour)
≈ 6,000
≈ 21,000
≈ 54,000
EHRMamba
Vendor: EHRMamba authors
What it does: models long structured record sequences efficiently, which matters when a patient history runs to thousands of events.
Task-dependent; evaluated on EHR prediction tasks using MIMIC-IV.
Requirement
Minimum
Medium
High
GPU type
RTX 3060
RTX 4090
A100 80 GB
VRAM
12 GB
24 GB
80 GB
vCPUs
6
12
24
RAM
16 GB
32 GB
64 GB
Server
1× RTX 3060 12 GB
1× RTX 4090 24 GB
1× A100 SXM 80 GB
Rate (documents/hour)
≈ 6,000
≈ 21,000
≈ 54,000
Choosing between them
Your data model decides the shortlist: OMOP or FHIR event streams suit the timeline foundation models directly. Expect local training rather than out-of-the-box use, and expect the honest comparison to be against your existing gradient-boosted baseline, which is what we benchmark against. Some weights in this group are released on request, which we handle for you.
Accuracy figures above are those the producers and independent evaluations report, on their own test sets. They are a shortlist tool, not a prediction of what you will see. At the start of a project we run a short proof of concept on a sample of your own data, which replaces them with real figures — so the cost and the schedule for the full engagement are known before anything is committed.
Send a representative sample, your expected volume and your latency target for ehr outcome prediction. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster — with the cost per unit of work known before you commit.