EDGE: KNOWLEDGE-DRIVEN NEW DRUG RECOM-MENDATION

Abstract

Drug recommendation assists doctors in prescribing personalized medications to patients based on their health conditions. Existing drug recommendation solutions adopt the supervised multi-label classification setup and only work with existing drugs with sufficient prescription data from many patients. However, newly approved drugs do not have much historical prescription data and cannot leverage existing drug recommendation methods. To address this, we formulate the new drug recommendation as a few-shot learning problem. Yet, directly applying existing few-shot learning algorithms faces two challenges: (1) complex relations among diseases and drugs and (2) numerous false-negative patients who were eligible but did not yet use the new drugs. To tackle these challenges, we propose EDGE, which can quickly adapt to the recommendation for a new drug with limited prescription data from a few support patients. EDGE maintains a drugdependent multi-phenotype few-shot learner to bridge the gap between existing and new drugs. Specifically, EDGE leverages the drug ontology to link new drugs to existing drugs with similar treatment effects and learns ontology-based drug representations. Such drug representations are used to customize the metric space of the phenotype-driven patient representations, which are composed of a set of phenotypes capturing complex patient health status. Lastly, EDGE eliminates the false-negative supervision signal using an external drug-disease knowledge base. We evaluate EDGE on two real-world datasets: the public EHR data (MIMIC-IV) and private industrial claims data. Results show that EDGE achieves 7.3% improvement on the ROC-AUC score over the best baseline.

1. INTRODUCTION

With the wide adoption of electronic health records (EHR) and the advance of deep learning models, we have seen great opportunities in assisting clinical decisions with deep learning models to improve resource utilization, healthcare quality, and patient safety (Xiao et al., 2018) . Drug recommendation is one of the essential applications which aims at assisting doctors in recommending personalized medications to patients based on their health conditions. Existing drug recommendation methods typically formulate it as a supervised multi-label classification problem (Zhang et al., 2017; Zitnik et al., 2018; Shang et al., 2019b; Yang et al., 2021; Rui Wu & Wu., 2022; Tan et al., 2022b) . They often train on massive prescription data to learn patient representations and use the learned representations to predict medications (i.e., labels). However, in reality, new drugs come to the market all the time. For example, U.S. Food and Drug Administration (FDA) approves a wide range of new drugs every year (FDA, 2022) . Most of these newly approved drugs do not have much historical data to support model training (Blass, 2021) . Even if sufficient prescription data for new drugs exists, existing models must be periodically re-trained or updated to recommend new drugs, which is expensive and complex. As a result, existing drug recommendation methods can only recommend the same set of drugs seen during training and are no longer applicable when new drugs appear. To address this, we formulate the recommendation of new drugs as a few-shot classification problem. Given a new drug with limited prescription data from a few support patients (e.g., from clinical trials (Duijnhoven et al., 2013) ), the model should quickly adapt to the recommendation for this drug. Meta-learning approaches have been widely used in such problems by learning how to quickly adapt the classifier to a new label unseen during training, given only a few support examples (Finn et al., 2017; Snell et al., 2017) . However, most prior meta-learning works focus on vision or language-related tasks. In the new drug recommendation, applying existing meta-learning algorithms faces the following challenges. (1) Complex relations among diseases and drugs: diseases and medicines can have inherent and higher order relations. Deciding whether to prescribe a drug to a specific patient depends on many factors, such as disease progression, comorbidities, ongoing treatments, individual drug response, and drug side effects. General meta-learning algorithms do not explicitly capture such dependencies. (2) Numerous false-negative patients: many drugs can treat the same disease, but usually, only one of them is prescribed. For any given drug, there exist many falsenegative patients who were eligible but did not yet use the new drug (e.g., due to drug availability, doctor's preference, or insurance coverage). The number of false-negative supervision signals will substantially confuse the model learning, especially in the few-shot learning setting. To address these challenges, we introduce EDGE, a drug-dependent multi-phenotype few-shot learner to quickly adapt to the recommendation for a new drug with limited support patients. Specifically, since drugs within the same category often have similar treatment effects, EDGE utilizes the drug ontology for drug representation learning to link new drugs with existing drugs. Further, EDGE learns multi-phenotype patient representations to capture the complex patient health status from different aspects such as chronic diseases, current symptoms, and ongoing treatments. Given a new drug with a few support patients, EDGE makes recommendations by performing a drug-dependent phenotype-level comparison between representations of query patients and corresponding support prototypes. Lastly, to reduce the false-negative supervision signal, EDGE leverages the MEDI (Wei et al., 2013) drug-disease knowledge base to guide the negative sampling process. The main contributions of this work include: • To our best knowledge, this is the first work formulating the task of new drug recommendation; • We propose a meta-learning framework EDGE to solve this problem by considering complicated relations among diseases and drugs, and eliminating numerous false-negative patients. • We conduct extensive experiments on the public EHR data MIMIC-IV (Johnson et al., 2020) and private industrial claims data. Results show that our approach achieves 5.6% over ROC-AUC, 6.3% over Precision@100, and 5.5% over Recall@100 when providing recommended patient lists for new drugs. We also include detailed analyses and ablation studies to show the effectiveness of multi-phenotype patient representation, drug-dependent patient distance, and knowledge-guided negative sampling.

2. PROBLEM FORMULATION AND PRELIMINARIES

Denote the set of all drugs as M; the goal of drug recommendation is to prescribe drugs in M that are suitable for a patient with a record v = [c 1 , . . . , c V ], which consists of a list of diseases (and procedures)foot_0 , and V is the total number of diseases and procedures in the record v. Prior works (Zhang et al., 2017; Shang et al., 2019b; Yang et al., 2021; Tan et al., 2022b) formulate drug recommendation as a multi-label classification problem by generating a multi-hot output of size |M|. However, this formulation assumes that the drug label space M remains unchanged after training and is not applicable when new drugs appear. Thus, we propose an alternative formulation for the new drug recommendation as follows. Assume the entire drug set M is partitioned into a set of existing drugs M old and a set of new drugs M new , where M old ∩ M new = ∅. Each existing drug m i ∈ M old has sufficient patients using the drug m i (e.g., from EHR data). Each new drug m t ∈ M new is associated with a small support set St = {vj} Ns j=1 consisting of patients using the drug m t (e.g., from clinical trials), and an unlabeled query patient set Qt = {vj} Nq j=1 , where N s and N q are the number of patients in the support and query sets, respectively. The goal of new drug recommendation is to train a model f ϕ (•) parameterized by ϕ on existing drugs M old , such that it can adapt to new drug m t ∈ M new given the small support set S t , and make correct recommendation on the query set Q t . Our work is inspired by the prototypical network (Snell et al., 2017) , which learns a representation model f ϕ (•) such that patients using a specific drug will cluster around a prototype representation. Recommendation can then be performed by computing the distance to the prototype. To equip the



To reduce clutter, we use a unified notation for both diseases and procedures. Since we focus on recordlevel prediction, "patient" and "record" are used interchangeably.

