2c:null 2d:["$","title",null,{"children":"Delphi-2M in the browser | Praful Bhawsar | Observable"}] 32:T5a6,getDelphiLogits = async ( diseaseTokens, ages, logitsIndex, withDims = false, withWebGPU = false ) => { const onnxSession = withWebGPU ? sessionWithWebGPU : session; // Only works for a batch size of 1! const idxTensor = new onnxRuntime.Tensor( "int64", new BigInt64Array(diseaseTokens.map((x) => BigInt(x))), [1, diseaseTokens.length] ); const ageTensor = new onnxRuntime.Tensor("float32", new Float32Array(ages), [ 1, ages.length ]); const feeds = { idx: idxTensor, age: ageTensor }; const results = await onnxSession.run(feeds); const { data: logits, dims: logitsShape } = results.logits; // const batchSize = logitsShape[0]; // Assuming batchSize always 1 for now, index calculation below will need adjustment otherwise const numEvents = logitsShape[1]; const vocabSize = logitsShape[2]; let logitsForEvent = []; // Return last event logits by default if (logitsIndex === -1) { // Return logits for all events logitsForEvent = logits; } else { let eventLogitsIndex = (numEvents - 1) * vocabSize; if (logitsIndex < logitsForEvent.length / vocabSize) { eventLogitsIndex = logitsIndex * vocabSize; } logitsForEvent = logits.slice( eventLogitsIndex, eventLogitsIndex + vocabSize ); } if (withDims) { return { logits: logitsForEvent, dims: logitsShape }; } else { return logitsForEvent; } }33:Td1a,async function generate(idx, age, options = {}) { const { maxNewTokens = 100, maxAge = 85 * 365.25, noRepeat = true, terminationTokens = [1269], ignoreTokens = [], outputName = null } = options; const maskTime = -10000; const finalMaxTokens = maxNewTokens === -1 ? 128 : maxNewTokens; let currentIdx = [...idx]; let currentAge = [...age]; for (let step = 0; step < finalMaxTokens; step++) { const logits = await getDelphiLogits(currentIdx, currentAge); for (const ignoreToken of ignoreTokens) { if (ignoreToken < logits.length) { logits[ignoreToken] = -Infinity; } } if (noRepeat) { const fill = currentIdx.map((token) => (token === 1 ? 0 : token)); for (const token of fill) { if (token > 0 && token < logits.length) { logits[token] = -Infinity; } } } const expLogits = logits.map((x) => Math.exp(-x)); const randomSamples = logits.map(() => Math.random()); const tSamples = expLogits.map((expLogit, i) => Math.max(0, Math.min(365 * 80, -expLogit * Math.log(randomSamples[i]))) ); let minTime = Infinity; let minIndex = 0; for (let i = 0; i < tSamples.length; i++) { if (tSamples[i] < minTime) { minTime = tSamples[i]; minIndex = i; } } const ageNext = currentAge[currentAge.length - 1] + minTime; currentIdx.push(minIndex); currentAge.push(ageNext); const hasTerminationToken = currentIdx.some((token) => terminationTokens.includes(token) ); const exceedsMaxAge = ageNext > maxAge; if (hasTerminationToken || exceedsMaxAge) { break; } } const pad = Array(currentIdx.length).fill(false); let terminationFound = false; let terminationCount = 0; for (let i = 0; i < currentIdx.length; i++) { if (terminationTokens.includes(currentIdx[i])) { terminationFound = true; terminationCount++; } if (terminationFound && terminationCount > 1) { pad[i] = true; } if (currentAge[i] > maxAge) { pad[i] = true; } } const finalLogitsOutput = await getDelphiLogits( currentIdx, currentAge, -1, true ); for (let i = 0; i < pad.length; i++) { if (pad[i]) { currentIdx[i] = 0; currentAge[i] = maskTime; } } let processedLogits = []; const vocabSize = finalLogitsOutput.dims[2]; if (noRepeat) { for (let seqPos = 0; seqPos < currentIdx.length; seqPos++) { const startIdx = seqPos * vocabSize; const positionLogits = finalLogitsOutput.logits.slice( startIdx, startIdx + vocabSize ); const fill = currentIdx .slice(0, seqPos + 1) .map((token) => (token === 1 ? 0 : token)); for (const token of fill) { if (token > 0 && token < positionLogits.length) { positionLogits[token] = -Infinity; } } processedLogits.push(positionLogits); } } else { for (let seqPos = 0; seqPos < currentIdx.length; seqPos++) { const startIdx = seqPos * vocabSize; const positionLogits = finalLogitsOutput.logits.slice( startIdx, startIdx + vocabSize ); processedLogits.push(positionLogits); } } return { idx: currentIdx, age: currentAge, logits: processedLogits }; }34:[{"id":"6a654d1deac7e342b946c7ea21d2b10738c96f5d1cdae209f1ce362e1ef8677f7dc42820eea4bdea40aa0d3e03e5eb1a8a9ce11bdb768d1bea8df37aaf10988e","href":"https://static.observableusercontent.com/files/6a654d1deac7e342b946c7ea21d2b10738c96f5d1cdae209f1ce362e1ef8677f7dc42820eea4bdea40aa0d3e03e5eb1a8a9ce11bdb768d1bea8df37aaf10988e","name":"delphi.onnx","type":"application/octet-stream","lastModified":1758657460339.514,"size":9758087}] 30:{"id":"b154408d9b156ebe","slug":"delphi-2m-in-the-browser","title":"Delphi-2M in the browser","thumbnail":"27edf0eb04a616842ae2086251a7098981425cc454063829cba9d40b40b3c034","access_level":"public","create_time":"$D2025-09-22T15:54:06.519Z","update_time":"$D2025-10-01T13:35:10.170Z","view_time":null,"first_public_time":"$D2025-09-23T21:07:38.110Z","has_unpublished_edits":null,"fork":false,"stars":1,"starred":false,"owner":{"type":"team","id":"b0b4412d0985303e","login":"prafulb","name":"Praful Bhawsar","avatar_url":"https://avatars-next.observableusercontent.com/765f5b661080e697f1ac35466b8e2cef535a03897b0dc7163c9b13484530ac2a","tier":"starter_2024"},"version":470,"publish_version":null,"role":null,"authored":false,"authors":[{"type":"individual","id":"49bdf3232a315bf5","login":"prafulb","name":"Praful Bhawsar","avatar_url":"https://avatars-next.observableusercontent.com/765f5b661080e697f1ac35466b8e2cef535a03897b0dc7163c9b13484530ac2a","tier":null}],"members":[{"type":"individual","id":"49bdf3232a315bf5","login":"prafulb","name":"Praful Bhawsar","avatar_url":"https://avatars-next.observableusercontent.com/765f5b661080e697f1ac35466b8e2cef535a03897b0dc7163c9b13484530ac2a","tier":null,"role":"editor"},{"type":"team","id":"b0b4412d0985303e","login":"prafulb","name":"Praful Bhawsar","avatar_url":"https://avatars-next.observableusercontent.com/765f5b661080e697f1ac35466b8e2cef535a03897b0dc7163c9b13484530ac2a","tier":"starter_2024","role":"editor"}],"comments":[],"custom_thumbnail":false,"trashed":false,"lapsed":null,"subscribed":null,"forks":2,"collection":null,"type":"notebook","body":{"titling":"auto","title":"Delphi-2M in the browser","cells":[{"id":0,"value":"# Delphi-2M in the browser","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"","database":"$undefined"},{"id":239,"value":"Original Paper: [Learning the natural history of human disease with generative transformers](https://www.nature.com/articles/s41586-025-09529-3)","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"","database":"$undefined"},{"id":24,"value":"exampleHealthTrajectory = [\n [\"Male\", 0],\n [\"B01 Varicella [chickenpox]\", 2],\n [\"L20 Atopic dermatitis\", 3],\n [\"No event\", 5],\n [\"No event\", 10],\n [\"No event\", 15],\n [\"No event\", 20],\n [\"G43 Migraine\", 20],\n [\"E73 Lactose intolerance\", 21],\n [\"B27 Infectious mononucleosis\", 22],\n [\"No event\", 25],\n [\"J11 Influenza, virus not identified\", 28],\n [\"No event\", 30],\n [\"No event\", 35],\n [\"No event\", 40],\n [\"Smoking low\", 41],\n [\"BMI mid\", 41],\n [\"Alcohol low\", 41],\n [\"No event\", 42]\n]","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":295,"value":"### Trajectory Generated by Delphi-2M:","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"","database":"$undefined"},{"id":141,"value":"viewof table = Inputs.table(generatedTrajectory, {\n columns: [\"Age\", \"Event\", \"Type\"],\n width: { Age: 20, Event: 300, Type: 50 },\n align: { Age: \"left\", Event: \"left\" },\n maxWidth: 900,\n select: false\n})","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":92,"value":"logits = await generate(events, ages)","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":119,"value":"logitsOut = {\n const labelNames = delphiLabels.map((row) => row[\"name\"]);\n const labelChapters = delphiLabels.map((row) => row[\"ICD-10 Chapter\"]);\n return logits.logits.map((logit, i) => {\n return { name: labelNames[i], chapter: labelChapters[i], logit };\n });\n}","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":95,"value":"

","pinned":false,"hidden":false,"mode":"html","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"","database":"$undefined"},{"id":5,"value":"onnxRuntime = import(\n \"https://cdn.jsdelivr.net/npm/onnxruntime-web/dist/ort.min.mjs\"\n)","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":457,"value":"session = {\n // onnxRuntime.env.logLevel = \"verbose\";\n // onnxRuntime.env.debug = true;\n return onnxRuntime.InferenceSession.create(\n \"https://episphere.github.io/delphi-onnx/delphi.onnx\",\n {\n executionProviders: [\"wasm\"]\n }\n );\n}","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":7,"value":"sessionWithWebGPU = {\n session;\n // onnxRuntime.env.logLevel = \"verbose\";\n // onnxRuntime.env.debug = true;\n return onnxRuntime.InferenceSession.create(\n \"https://episphere.github.io/delphi-onnx/delphi.onnx\",\n {\n executionProviders: [\"webgpu\"]\n }\n );\n}","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":45,"value":"$32","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":316,"value":"$33","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"},{"id":195,"value":"getTrajectory = async (maxEvents = 1) => {\n const {\n idx: predictedEvents,\n age: predictedAges,\n logits\n } = await generate(events, ages);\n const trajectory = [];\n\n predictedEvents.forEach((event, i) => {\n const predictedEvent = delphiLabels.find(\n (entry) => entry.index == event\n ).name;\n const predictedAge = (predictedAges[i] / 365).toFixed(1);\n\n const obj = {};\n obj[\"Age\"] = predictedAge;\n obj[\"Event\"] = predictedEvent;\n obj[\"Type\"] =\n i <= exampleHealthTrajectory.length - 1 ? \"Input\" : \"Prediction\";\n trajectory.push(obj);\n });\n return trajectory;\n}","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","providerMetadata":"$undefined","attachments":"$undefined","output":"$undefined","database":"$undefined"}],"files":"$W34","stdlib":"1"}} 31:"$30:body"