Reading the clinician's mind: using machine learning to assess the assessors
Featuring Matt Metherell, Senior Programmer II at PHASTAR
Abstract: 'Reading the Clinician's Mind: Using Machine Learning to assess the assessors'.
We present an initial investigation into whether the input of the clinician can be predicted using machine learning approaches on the Adverse Event (AE) dataset. We address whether it is possible to predict if an AE was related to the investigational product or predict the dosing decisions by the clinician (interruptions, dose alterations etc) based purely on the Adverse Event dataset. Furthermore, we consider whether the inclusion of additional datasets such as demographic data make these predictions more accurate.
We will discuss the implications and applications of these predictions: If we are able to predict a clinicians dosing decision, what would the implications of that be? Can we query any decisions that we consider 'abnormal'? Certainly, we may uncover data issues and if we uncover systemic differences, perhaps by site or geographical region, we may want to consider treating the safety analysis differently for each case. We may even be able to predict which subjects are in danger of having their treatment discontinued and step in before that occurs.
Early indications are promising based on an analysis on freely available (but limited) AE data from the National Institute on Drug Abuse (NIH) Data Share Website (https://datashare.nida.nih.gov/). Initial training models were able to produce -65% accuracy for the drug relationship, demonstrating the viability of this idea. This approach has broader applicability across larger, more detailed clinical trial datasets where we may find certain diseases or bodily systems are more predictable than others.