In plain words
Drug reviews are a goldmine of unstructured medical information — but they almost never state plainly which disease the drug treats. This work predicts the disease (or disease type) directly from a drug's name and its review text.
Using the UCI ML Drug Review dataset — 215,063 reviews from Drugs.com — the authors compared classical machine learning (Random Forest, Multinomial Naive Bayes) against a two-layer LSTM with Google News word embeddings.
Random Forest won: accuracy 0.83, precision 0.89, recall 0.72, F1 0.75. The LSTM with Google News embeddings followed at 0.81 accuracy (F1 0.80), while Naive Bayes trailed. The authors position it as the first work to predict disease names from drug reviews. (Preprint — figures are the authors' reported results.)
How it works
- Start with scale215,063 real patient reviews from Drugs.com form the corpus.
- Turn text into featuresReviews become numerical features — bag-of-words for classical models, word embeddings for the LSTM.
- Race the modelsRandom Forest, Multinomial Naive Bayes and a two-layer LSTM each learn the review-to-disease mapping.
- Crown the winnerRandom Forest takes it at 0.83 accuracy, with the embedding-powered LSTM close behind.
Why it matters
Mapping drugs to the conditions they treat — automatically, from public text — supports pharmacovigilance, patient education and health-information systems in places where structured medical databases are incomplete or unaffordable.