Serendipity-A Machine-Learning Application for Mining Serendipitous Drug Usage From Social Media

Area of Science:

• Computational linguistics
• Pharmacology
• Artificial Intelligence

Background:

• Serendipitous drug usage, the unexpected therapeutic effect of a drug on an unindicated condition, has historically led to new drug discoveries.
• Patient-reported outcomes on social media offer a rich, albeit unstructured, source for identifying such serendipitous events.
• Computational identification of these events can accelerate the generation and validation of drug-repositioning hypotheses.

Purpose of the Study:

• To investigate the efficacy of deep neural network (DNN) models in mining patient-reported serendipitous drug usage from social media data.
• To enhance DNN models with contextual information, medical ontologies, and knowledge bases for improved accuracy.
• To compare the performance of these advanced DNNs against traditional machine learning algorithms.

Main Methods:

• Utilized word2vec for creating word-embedding features from WebMD drug reviews.
• Adapted and redesigned Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Convolutional LSTM (ConvLSTM) models.
• Integrated contextual information, information-filtering tools, medical ontology, and knowledge into the DNN architectures.
• Trained and evaluated models on a gold-standard dataset with a low prevalence (2.8%) of serendipitous drug usage sentences.
• Compared DNN performance against Support Vector Machine (SVM), Random Forest, and AdaBoost.M1 algorithms.

Main Results:

• Contextual information significantly reduced the false-positive rate in DNN models.
• In highly imbalanced datasets, DNNs did not outperform traditional models when using n-gram and context features.
• DNNs demonstrated a superior ability to leverage word embeddings for feature construction, highlighting their potential.
• A web-based application was developed, integrating natural language processing (NLP) and machine learning (ML) for mining social media.

Conclusions:

• Deep neural networks show promise for mining serendipitous drug usage from social media, aiding drug discovery.
• Integrating contextual information and word embeddings enhances the capability of these models.
• Further investigation into DNNs for this task is warranted due to their advanced feature extraction capabilities.
• The developed application provides a tool for researchers to explore social media for novel drug-use insights.