An intelligent machine learning model for real-time early detection of undesirable cancer events: AIM2REDUCE.

Abstract

1557 Background: Cancer and its treatment cause undesirable cancer events (UCEs). Automated warning systems could reduce the frequency and severity of UCEs by alerting the healthcare team and allocating preventative interventions. Most previous studies predicted single UCEs at the initiation of treatment. In AIM2REDUCE, we developed and evaluated a general-purpose system for predicting UCEs during outpatient systemic anti-cancer therapy. Methods: Each time a patient receives treatment, AIM2REDUCE applies machine learning to the preceding data in the electronic medical record (EMR) to predict future UCEs. We identified patients treated for aerodigestive cancers from the EMR at Princess Margaret Cancer Centre, who were randomly split into development, validation, and test cohorts. Features included cancer diagnosis, treatment sessions with doses, laboratory tests, and patient-reported symptoms. UCEs are listed in the Table. We trained LASSO regression and random forests models in the training cohort and tuned hyperparameters in the validation cohort using Bayesian optimization. We evaluated performance across discrimination, calibration, and net benefit in the test cohort. Results: The cohort included 5,760 patients who received 175,565 treatment sessions, with 13,612,746 unique data points across 102 features. Of these patients, 2,352 (40.8%) were female, the median age was 64.0 years (interquartile range 14.0), the most common diagnoses were lung cancer (2,071, 36.0%) and pancreatic cancer (926, 16.1%), and the most common treatment regimens were weekly gemcitabine (433, 7.5%) and maintenance pemetrexed (417, 7.2%). The Table shows the performance of AIM2REDUCE. Conclusions: We demonstrate that longitudinal machine learning systems trained using EMR data can accurately predict a wide range of UCEs. Based on these results, automated warning systems should be implemented and evaluated in real-time clinical practice. [Table: see text]