1
Using indication embeddings to represent patient health for drug safety studies Rachel Melamed Biomedical Data Science @University of Chicago & Biology @ UMass Lowell [email protected] | @RDMelamed Goal: high-throughput drug safety studies Randomized trials : low-throughput but unbiased 1) From data select people.. ✄ ☤ ☤ ✄ ☤ ✄ ☤ ☤ ☤ 1) Enroll cohort 2) Randomize treatment 3) Compare experimental groups Cohort studies : reuse health data to emulate randomized trial. Drug safety Does taking this drug change your risk of some health outcome? Exposure Cancer ☤ Health data X- ray asthma statin arthritis ✄ ☤ 5,000 Precriptions 10,000 Diagnosis codes 20,000 Procedure codes …taking treatment drug …or comparator drug Can we match without expert design? Creating indication embeddings Evaluating embeddings The challenge: confounding Embeddings identify comparator drugs Match with embeddings age aspirin P(treated | age,..) Propensity score match HEALTH age aspirin ☤ diabetes Evaluate with propensity score: P(treated) in treated cohort P(treated) in comp. P(treated | …) 2) Match to emulate randomization ☤ High-throughput cohort studies Currently, cohort study design relies on domain experts: never often 25 75 age aspirin Expert Task 1 Find suitable comparator drug Expert Task 2 Design matching— identify confounders The solution: matching age Match on confounders ! "#$%"$& %'$, %)!*#*+, … ) Insulin resistance insulin, Type 2 diabetes amoxicillin xanax …other Rx, Dx, Px How to match on 30,000+ dimensional, sparse, uninformative vectors? Instead map them to small, meaningful embeddings Indication embedding Drug embedding Training task: Predict drug Simple neural network 150 million patient histories ☤ blood lab gout statin diabetes ✄ New Rx metformin History event New Rx ( , ) ( , ) Create training examples ( , ) ( , ) ✄ Embeddings relate codes to health needs ☤ Drug embedding = drugs given in most similar health contexts Indication embedding = health context for prescription of a new drug Map each event to 50- dimensional vector For each drug, performance of embedding distance to predict indications. Overall ROCAUC = .82 auc Dot-products between antidepressants and selected closest diagnoses. tricyclic SSRI SNRI anticonvulsants antipsychotics Drugs with closest embedding dot product are more comparable, as measured by AUC Drugs with same therapeutic use as carbamazepine: primarily anticonvulsant, off-label for bipolar. aripiprazole asenapine carbamazepine chlorpromazine clozapine divalproex_sodium fluphenazine gabapentin haloperidol haloperidol_lactate lacosamide lamotrigine levetiracetam lithium_carbonate lurasidone Olanzapine ✄ ☤ ☤ ✄ ☤ ✄ ☤ ☤ ☤ ROC AUC Expert Task 1 Therapeutic class oxcarbazepine paliperidone perphenazine primidone prochlorperazine prochlorperazine_maleate quetiapine_fumarate risperidone thioridazine thiothixene tiagabine topiramate trifluoperazine valproic_acid ziprasidone zonisamide 2003 2013 25 75 age year Step 1 : Coarsened exact matching by age, gender, year, number of Rx Step 2 : Encode histories à small dense vectors Step 3 : Mahalanobis match on health summaries within bins Indication embedding (R V x E ) ✄ Weighted average (upweight recent history) ✄ ☤ Embedding matching Expert Task 2 ! ☤ Embedding can match for key confounders Propensity score matching Matching people on bupropion to trazodone is complicated by alternate indication of bupropion for smoking cessation. Each point is one person on bupropion, trazodone, or varenicline. Health summary vectors Embedding better matches nonsmokers to nonsmokers 0 0 1 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 .2 1 0 .1 .8 Then do simple nearest- neighbor matching
