A Machine Learning Model to Improve Immune Checkpoint Inhibitor Utility

Downloadable Content

Read in Browser Download

Item Description

  • Since the approval of Ipilimumab in 2011 use of immune checkpoint inhibitors (ICIs) has become a mainstay of cancer therapy with the potential to dramatically improve outcomes for even patients with advanced disease. Despite the dramatic potential for these innovative therapies, ICIs come with side effects which can be extensive. Immune related adverse events (irAEs) have been reported in between 12-79% of patients taking ICIs and contribute to increased morbidity and healthcare utilization by cancer patients. Our team created a machine learning model to proactively identify those patients most at risk of irAEs and target them for intervention in the outpatient setting before they require hospital admission or emergency care. Our light gradient boosted model was trained using data from the 3962 adult patients who received CTLA4 and/or PD-1/PD-L1 inhibitors at Duke University Hospitals between April 2016-July 2021. We predicted the risk of any ED visit or hospital admission in the next two weeks for the first 6 months of therapy. Inputs for the model consisted of over 150 features including lab values, vitals, medication type and duration, demographics, comorbidities, visit reasons, and encounter level data. Retrospective data showed our outcomes occurred in 35.5% of our cohort with a majority of events occurring in the initial weeks of therapy. 12.2% of patients with outcomes were seen that same day in outpatient clinics. Model performance yielded an AUC of 0.76. Factors with high importance include length of time on immunotherapy, number of prior encounters in the Duke system, sodium, albumin, TSH, and pulse. The model will be validated by targeted clinicians before implementation in outpatient GU and thoracic cancer centers. Our model generates a risk score which easily encapsulates the likelihood of an ED visit or hospital admission in the next 2 weeks for patients receiving ICIs. Allowing providers to proactively intervene on patients at high risk of an adverse event will increase the utility and minimize risk of this important field of cancer therapy.
Date created
  • 0000-0002-5753-4113
Research type
Study program
Research Location
In Collection:

QR Code