Integrating machine learning to improve kidney disease risk assessment for patients on cancer therapy - ABSTRACT As evolving therapies enable a longer lifespan for patients with cancer, risk for kidney function decline, either pre-existing or developed as a complication of therapy, requires increasing attention. Many factors that render serum creatinine (sCr) or cystatin-C based kidney function assessments unreliable—including sarcopenia for the first or inflammation for the latter—are enriched in this population. Furthermore, widely used molecularly targeted cancer medications may inhibit tubular creatinine secretion and raise serum sCr; little is known about the long-term effects of these medications that once initiated, require long-term, daily intake. Therefore, uncertainty about the risk for clinically significant kidney function decline among patients on molecularly targeted cancer therapy is complicated by uncertainty about the performance of current kidney function measures in this population. Under mentorship from experts in care of patients with cancer and kidney disease, and in bioinformatics, Dr. Ziolkowski will address critical data gaps in assessment of kidney function among patients with cancer. She will take coursework in bioinformatics and apply machine learning methodology to harness the rich health data available for a person with cancer. In a retrospective cohort of patients treated at Stanford Health care, our team will determine the two-year incidence rate and risk factors for progressive chronic kidney disease (CKD) for patients on cancer therapies associated with an acute rise in sCr at drug initiation, with the hypothesis that long term exposure to these medications may contribute to progressive CKD. We will obtain sCr, cystatin C, and kidney biomarkers of tubular injury pre- and post- drug start to characterize the acute effects of drug initiation. We will develop a risk assessment model using machine learning for patients on molecularly-targeted cancer therapies to predict risk of progressive CKD, with the hypothesis that integrating multi-dimensional individual-level data in the electronic medical record will have greater prognostic yield for progressive CKD than estimated glomerular filtration rate cut points. We anticipate incorporation of computed tomography (CT) measures of muscle mass and kidney size improve our risk stratification. We will externally validate the prediction model at Mount Sinai Health System. A risk stratification model could identify patients at risk for progressive CKD, providing opportunities for appropriate counseling, drug dosing, avoidance of nephrotoxins, and treatment with renal-protective medications while those at lower risk could have expanded treatment options. Additional training in risk assessment methodologies and machine learning will enable Dr Ziolkowski to become an independent investigator, applying the latest bioinformatics methodologies to improve kidney function outcomes among patients with cancer. She will apply for an investigator led (R01) grant to determine clinical utility of our risk assessment model, using simulated and real-world prospective data, and broaden the risk assessment methodology to other cancer populations.
