JavaScript SDK for Delphi-2M (+ ONNX Export scripts)

Reproducible evaluation

The evaluation compares OriginalModel.pt with delphi.onnx through the actual ONNX Runtime Web JavaScript path. It uses a fixed, seeded cohort from the upstream synthetic validation data, evaluates logits immediately before random sampling, checks the corresponding masked event distributions, and benchmarks PyTorch CPU, browser Wasm, and browser WebGPU on the development machine.

• Publication-ready results
• Evaluation design and reproduction instructions
• Machine-readable measurements

Example Usage

Trajectory Generation:

let eventsList = [
    {
        "event": "Male",
        "age": 0
    },
    {
        "event": "B01 Varicella [chickenpox]",
        "age": 2
    },
    {
        "event": "L20 Atopic dermatitis",
        "age": 3
    },
    {
        "event": "No event",
        "age": 5
    },
    {
        "event": "No event",
        "age": 10
    },
    {
        "event": "No event",
        "age": 15
    },
    {
        "event": "No event",
        "age": 20
    },
    {
        "event": "G43 Migraine",
        "age": 20
    },
    {
        "event": "E73 Lactose intolerance",
        "age": 21
    },
    {
        "event": "B27 Infectious mononucleosis",
        "age": 22
    },
    {
        "event": "No event",
        "age": 25
    },
    {
        "event": "J11 Influenza, virus not identified",
        "age": 28
    },
    {
        "event": "No event",
        "age": 30
    },
    {
        "event": "No event",
        "age": 35
    },
    {
        "event": "No event",
        "age": 40
    },
    {
        "event": "Smoking low",
        "age": 41
    },
    {
        "event": "BMI mid",
        "age": 41
    },
    {
        "event": "Alcohol low",
        "age": 41
    },
    {
        "event": "No event",
        "age": 42
    }
]

const { generateTrajectory } = await import("https://episphere.github.io/delphi-onnx/delphiSDK.js")
await generateTrajectory({eventsList, seed: 42})

Embeddings:

let eventsList = [
    {
        "event": "Male",
        "age": 0
    },
    {
        "event": "B01 Varicella [chickenpox]",
        "age": 2
    },
    {
        "event": "L20 Atopic dermatitis",
        "age": 3
    },
    {
        "event": "No event",
        "age": 5
    },
    {
        "event": "No event",
        "age": 10
    },
    {
        "event": "No event",
        "age": 15
    },
    {
        "event": "No event",
        "age": 20
    },
    {
        "event": "G43 Migraine",
        "age": 20
    },
    {
        "event": "E73 Lactose intolerance",
        "age": 21
    },
    {
        "event": "B27 Infectious mononucleosis",
        "age": 22
    },
    {
        "event": "No event",
        "age": 25
    },
    {
        "event": "J11 Influenza, virus not identified",
        "age": 28
    },
    {
        "event": "No event",
        "age": 30
    },
    {
        "event": "No event",
        "age": 35
    },
    {
        "event": "No event",
        "age": 40
    },
    {
        "event": "Smoking low",
        "age": 41
    },
    {
        "event": "BMI mid",
        "age": 41
    },
    {
        "event": "Alcohol low",
        "age": 41
    },
    {
        "event": "No event",
        "age": 42
    }
]

const { getEmbeddings } = await import("https://episphere.github.io/delphi-onnx/delphiSDK.js")
await getEmbeddings({eventsList, pooling: 'mean'}) // pooling could be mean, max, or last. If not specified, embeddings for all events will be returned.

Export script:

For the predictive model:

python export_onnx.py --checkpoint ckpt.pt --output model.onnx

For the embeddings only model:

python export_onnx.py --checkpoint ckpt.pt --output embeddingsModel.onnx --embeddings-only