Leave No Patient Behind: Enhancing Medication Recommendation for Rare Disease Patients

TL;DR

RAREMed addresses fairness gaps in medication recommendation for rare diseases by leveraging pretrain-finetune learning with two self-supervised tasks (Sequence Matching Prediction and Self Reconstruction) on a unified disease-procedure sequence. A transformer encoder learns robust patient representations, enabling accurate multi-label medication predictions while accounting for DDI safety. Evaluations on MIMIC-III and MIMIC-IV show improved accuracy for rare disease patients and reduced group-based performance gaps, validating the approach's practicality. The work contributes a fairness-focused medication recommender and releases code for reproducibility.

Abstract

Medication recommendation systems have gained significant attention in healthcare as a means of providing tailored and effective drug combinations based on patients' clinical information. However, existing approaches often suffer from fairness issues, as recommendations tend to be more accurate for patients with common diseases compared to those with rare conditions. In this paper, we propose a novel model called Robust and Accurate REcommendations for Medication (RAREMed), which leverages the pretrain-finetune learning paradigm to enhance accuracy for rare diseases. RAREMed employs a transformer encoder with a unified input sequence approach to capture complex relationships among disease and procedure codes. Additionally, it introduces two self-supervised pre-training tasks, namely Sequence Matching Prediction (SMP) and Self Reconstruction (SR), to learn specialized medication needs and interrelations among clinical codes. Experimental results on two real-world datasets demonstrate that RAREMed provides accurate drug sets for both rare and common disease patients, thereby mitigating unfairness in medication recommendation systems.

Figures (5)

• Figure 1: (a) The long-tail distribution of disease codes in the MIMIC-III mimic-iii and MIMIC-IV mimic-iv datasets. (b) Patients in MIMIC-IV are divided into five equal-sized groups based on the ranking of their least common disease codes in terms of popularity. The subgraph illustrates the recommendation accuracy, as measured by the Jaccard index, of SOTA methods for both the commonest and the rarest code groups.
• Figure 2: We categorize patients into 13 distinct groups based on the frequency of their least common disease codes. Subplots (a), (b), (c), and (d) depict the relationship between the average disease count, procedure count, drug count, and drug popularity, respectively, w.r.t. the rarest disease popularity.
• Figure 3: The illustration of our proposed RAREMed. We first model the patient clinical information with three embedding layers and a transformer encoder to get the patient representation. Then we pre-train our model on Sequence Matching Prediction task and Self Reconstruction task. Finally, the model is fine-tuned on medication recommendation task to get the recommended medication combination.
• Figure 4: (a) Jaccard coefficients on various patient groups. (b) Standard deviation of Jaccard coefficients across groups. (c) and (d) The improvement ratio of Jaccard across groups w.r.t. the number of pre-training epochs. The bold number indicates the selected hyper-parameter.
• Figure 5: Hyperparameter effects on model performance.