Drug recommendation ranking for personalized medicine using outcomes of retrospective cancer patients

Abstract

The vast amount of now available genomic, biomolecular, and clinical patient data may revolutionize personalized medicine. Yet, this availability poses significant challenges to artificial intelligence as the major obstacle is that retrospective patients had only one treatment, and there is no way to determine what would have happened if they had other treatments. In this paper, we propose a novel way to use historical retrospective patient data to build systems that rank different treatments according to the possible benefit for individuals with specific tumor characteristics. This novel approach – our Drug Rankers over Historical Cases (DRHC) – solves the major obstacle for using retrospective patients by proposing a way to rank treatments in training and by introducing two novel measures (average ranking score and soft average ranking score) to evaluate these novel DRHCs Additionally, we categorize these novel DRHC methods into predictor-based and selector-based families. The latter leverages our novel measures, making it more innovative. Both families were implemented with various learning algorithms. Experiments demonstrate that the selector-based family outperforms the predictor-based family in terms of Average Ranking Score (ASR) and Average Soft Ranking Score (ASRS), showcasing superior drug ranking capabilities. The selector-based family achieves a maximum ASR of 0.642 and ASRS of 0.641, compared to 0.638 ASR and 0.634 ASRS for the predictor-based family.