Adverse Effects Image Fingerprinting using CNN Deep Learning as a Means for Individual Case Safety Report Classification.
Abstract: Objective We investigated the use of image recognition using convolutional neural networks (CNNs) to classify individual case safety reports (ICSRs) as serious or non-serious, demonstrating that it is possible to precede the well-established challenge of data extraction for prediction model development.
Method Using deep learning CNN, we extract data from images to predict case-level seriousness from incoming ICSR source documents.
Results: Our results demonstrate success in training a CNN to classify ICSRs using only image representations of source documents as input with an accuracy of 86% when compared against the ground truth.
Conclusion: Image recognition applied to PV data will allow us to precede the challenge of multiple document formats, transcription and extraction of data entities from source documents, allowing us to classify source documents upon intake.
This methodological breakthrough is significant as it increases the speed at which models may be trained to understand domain concepts, within PV and beyond, using image representations of documents. The authors believe this research may be applicable across industries which are not yet digitally ready enough to undertake a large-scale AI/ML initiative.
Grigoriy Serbarinov is a Senior Data Science Manager of Pharmacovigilance Innovation at Bristol-Myers Squibb. He brings with him more than 5 years of Machine Learning and Data Science experience and over 20 years working with Data Analytics and Big Data. He holds both a Master’s and Bachelor’s Degree in Physics and has demonstrated success in applying AI/ML in the financial sector before bringing his talents into the pharmaceutical industry. He’s interested in accelerating the pharmaceutical industry with the latest deep learning technologies.