[Paper Review] Predicting Discharge Medications at Admission Time Based on Deep Learning

Yang Yuan, Pengtao Xie+5 more|arXiv (Cornell University)|Nov 4, 2017

Machine Learning in Healthcare32 references3 citations

TL;DR

This paper proposes a deep learning approach using a Convolutional Neural Network (CNN) to predict discharge medications from admission notes, leveraging semantic representations and pharmacological correlations among medications. The method achieves a 20% improvement in macro-averaged F1 score over the best baseline, particularly excelling on infrequent medications due to its ability to capture medication correlations.

ABSTRACT

Predicting discharge medications right after a patient being admitted is an important clinical decision, which provides physicians with guidance on what type of medication regimen to plan for and what possible changes on initial medication may occur during an inpatient stay. It also facilitates medication reconciliation process with easy detection of medication discrepancy at discharge time to improve patient safety. However, since the information available upon admission is limited and patients' condition may evolve during an inpatient stay, these predictions could be a difficult decision for physicians to make. In this work, we investigate how to leverage deep learning technologies to assist physicians in predicting discharge medications based on information documented in the admission note. We build a convolutional neural network which takes an admission note as input and predicts the medications placed on the patient at discharge time. Our method is able to distill semantic patterns from unstructured and noisy texts, and is capable of capturing the pharmacological correlations among medications. We evaluate our method on 25K patient visits and compare with 4 strong baselines. Our methods demonstrate a 20% increase in macro-averaged F1 score than the best baseline.

Motivation & Objective

• To address the clinical challenge of predicting discharge medications at admission time, when patient information is limited and evolving.
• To improve patient safety by enabling early detection of medication discrepancies through predictive reconciliation.
• To develop a data-driven method that learns from unstructured clinical notes and captures complex pharmacological relationships among medications.
• To outperform existing baselines in predicting multiple discharge medications, especially for low-frequency drugs.
• To demonstrate the feasibility of using deep learning to support clinical decision-making in treatment planning using only admission data.

Proposed method

• A Convolutional Neural Network (CNN) is trained on raw admission notes to predict discharge medications, learning semantic representations from unstructured text.
• The model uses word embeddings to encode clinical text and convolutional filters to extract local patterns at multiple granularities.
• Shared latent factors in the dense layers enable the model to capture pharmacological correlations among medications, improving predictions for rare drugs.
• The architecture integrates admission medications, past medical history, chief complaint, and other clinical information as input features.
• Model training employs a multi-label classification framework with sigmoid activation to predict multiple medications simultaneously.
• Visualizations of word embeddings, convolutional filters, and dense layers are used to interpret learned representations and validate semantic understanding.

Experimental results

Research questions

• RQ1 Can a deep learning model accurately predict discharge medications using only admission notes, despite limited and noisy clinical information?
• RQ2 How does the model’s ability to capture medication correlations improve prediction performance, especially for infrequent medications?
• RQ3 To what extent does the CNN outperform traditional machine learning baselines in predicting multiple discharge medications?
• RQ4 How do semantic representations from raw text contribute to improved prediction accuracy compared to simpler feature-based models?
• RQ5 What role do pharmacological correlations play in enhancing model generalization across rare and common medications?

Key findings

• The CNN model achieves a 20% relative improvement in macro-averaged F1 score compared to the best baseline, demonstrating significant gains in overall prediction performance.
• The model shows greater improvement on infrequent medications such as atenolol and diltiazem, where performance gains are most pronounced due to effective correlation learning.
• For the two most frequent medications, metoprolol and furosemide, the CNN achieves F1 scores of 0.79 and 0.70, respectively, indicating strong performance on common drugs.
• The model’s micro-averaged F1 score is 0.63 precision at 0.70 recall, indicating balanced performance across the full medication set.
• Multilayer Perceptron (MLP) performs worst due to reliance solely on admission medications, highlighting the importance of integrating broader clinical context.
• The model’s superior performance on rare medications is attributed to its ability to leverage pharmacological correlations, compensating for data scarcity.