Abstract
In healthcare, the highest risk individuals for morbidity and mortality arerarely those with the greatest modifiable risk. By contrast, many machinelearning formulations implicitly attend to the highest risk individuals. Wefocus on this problem in point processes, a popular modeling technique for theanalysis of the temporal event sequences in electronic health records (EHR)data with applications in risk stratification and risk score systems. We showthat optimization of the log-likelihood function also gives disproportionateattention to high risk individuals and leads to poor prediction results for lowrisk individuals compared to ones at high risk. We characterize the problem andpropose an adjusted log-likelihood formulation as a new objective for pointprocesses. We demonstrate the benefits of our method in simulations and in EHRdata of patients admitted to the critical care unit for intracerebralhemorrhage.