[Paper Review] Automated Identification of Drug-Drug Interactions in Pediatric Congestive Heart Failure Patients

Daniel Miller|arXiv (Cornell University)|Feb 11, 2017

Heart Failure Treatment and Management3 references3 citations

TL;DR

This study presents an automated NLP pipeline to identify drug-drug interactions (DDIs) in pediatric congestive heart failure (CHF) patients by analyzing PubMed abstracts and electronic medical records. Using a Lasso classifier on word counts from 69,713 filtered abstracts, the method achieved high sensitivity (0.994) and positive predictive value (0.968), demonstrating the feasibility of scalable, data-driven DDI detection without clinical prior knowledge.

ABSTRACT

Congestive Heart Failure, or CHF, is a serious medical condition that can result in fluid buildup in the body as a result of a weak heart. When the heart can't pump enough blood to efficiently deliver nutrients and oxygen to the body, kidney function may be impaired, resulting in fluid retention. CHF patients require a broad drug regimen to maintain the delicate system balance, particularly between their heart and kidneys. These drugs include ACE inhibitors and Beta Blockers to control blood pressure, anticoagulants to prevent blood clots, and diuretics to reduce fluid overload. Many of these drugs may interact, and potential effects of these interactions must be weighed against their benefits. For this project, we consider a set of 44 drugs identified as specifically relevant for treating CHF by pediatric cardiologists at Lucile Packard Children's Hospital. This list was generated as part of our current work at the LPCH Heart Center. The goal of this project is to identify and evaluate potentially harmful drug-drug interactions (DDIs) within pediatric patients with Congestive Heart Failure. This identification will be done autonomously, so that it may continuously update by evaluating newly published literature.

Motivation & Objective

• To develop an automated system that identifies harmful drug-drug interactions (DDIs) in pediatric CHF patients using only published literature.
• To overcome data imbalance in DDI annotation by applying undersampling and regularization techniques.
• To evaluate model performance using clinical relevance metrics and future clinician feedback on predicted DDI time windows.
• To enable continuous, real-time DDI monitoring by integrating literature analysis with electronic medical record (MAR) data.
• To balance false positives and false negatives in clinical alerting systems to maintain sensitivity while minimizing noise.

Proposed method

• Extracted 69,713 abstracts from PubMed Central Open Access Subset containing at least one of 1781 drugs interacting with 44 pediatric CHF-relevant drugs.
• Used DrugBank as a gold standard for DDI labels and applied NLP techniques to map drug names and interaction types from abstracts.
• Implemented a supervised learning pipeline using word count features, with model training and validation split on a 70/15/15 ratio.
• Evaluated multiple models: raw and undersampled linear SVMs, raw and undersampled CNNs, and Lasso and logistic regression with word counts and GloVe embeddings.
• Optimized hyperparameters via cross-validation using AUC as the primary reward function, particularly for the Lasso model.
• Selected the final model based on validation set performance, with retraining on combined training and validation data before final test set evaluation.

Experimental results

Research questions

• RQ1 Can a machine learning model trained solely on PubMed abstracts detect clinically relevant drug-drug interactions in pediatric CHF patients with high sensitivity and precision?
• RQ2 How effective are different NLP-based feature representations—word counts, GloVe embeddings, and deep neural networks—on imbalanced DDI classification tasks?
• RQ3 Can L1 regularization (Lasso) outperform traditional models like SVM and CNN in identifying rare but critical DDIs in highly imbalanced datasets?
• RQ4 To what extent can automated literature analysis replace or augment expert-curated DDI databases like DrugBank?
• RQ5 How can the system be extended to generate time-specific DDI predictions by aligning abstract-derived interactions with actual medication administration records?

Key findings

• The Lasso classifier using word count features achieved the highest performance among all models, with 99.4% sensitivity and 96.8% positive predictive value (PPV) on the validation set.
• Undersampling improved model performance on imbalanced data, particularly for SVM and CNN models, but did not resolve the core issue of class imbalance.
• The CNN model trained on the 100 most frequent words showed promising results (73.3% sensitivity, 85.0% PPV) but underperformed compared to the Lasso model.
• GloVe embeddings did not outperform word counts; both logistic regression and Lasso with GloVe yielded lower AUC and lower sensitivity than the word count Lasso.
• The Lasso model demonstrated the most stable and effective performance, suggesting that feature selection is critical in high-dimensional, sparse NLP tasks for medical DDIs.
• The study confirms that automated, scalable DDI detection is feasible using only published literature and electronic health records, without requiring external clinical annotations.