Decision Support Tool Can Predict Patient Response to Treatment with Hypomethylating Agents
Researchers have developed the first machine-learning model that can predict with high accuracy MDS patient response or resistance to HMA treatment.
Hypomethylating agents (HMAs) can improve cytopenia and survival for myelodysplastic syndromes (MDS) patients, but only 30-40% of patients respond to treatment. Cleveland Clinic researchers have developed the first machine-learning model that can predict with high accuracy MDS patient response or resistance to HMA treatment.
Cleveland Clinic is a non-profit academic medical center. Advertising on our site helps support our mission. We do not endorse non-Cleveland Clinic products or services Policy
Having a predictive model can improve patient outcomes, decrease cost and toxicities, and steer treatment to alternative therapies when a therapeutic response is unlikely. “Most MDS patients who undergo HMA treatment for six months do not improve. It disrupts their lives and causes side effects, often with no benefit. We decided to find out if blood counts taken during the first three months of treatment would predict response at six months,” says Aziz Nazha, MD, a hematologist/oncologist and leader of the Cleveland Clinic Center for Clinical Artificial Intelligence, who developed the model with medical student Nathan Radakovich.
The findings were presented at the 2020 American Society of Hematology annual meeting.
The Center for Clinical Artificial Intelligence has developed other machine-learning-based decision support tools for hematologic cancers, including a geno-clinical model that can distinguish MDS from other myeloid malignancies; a predictive model for MDS patient survival which was cited as a major innovation by 35 media outlets; and an AML model that predicts patient survival.
The MDS model was developed and validated during 12 weeks of HMA (azacitidine or decitabine) treatment in a cohort of 414 patients who had their CBCs with differential monitored every one to two weeks during the first 12 weeks of therapy. A time series analysis of serial changes in blood count parameters using machine learning technology was used to develop the model, analogous to voice recognition algorithms, in which the sequence of words allows these algorithms to understand sentences. Similarly, monitoring changes in blood counts and the patterns of these changes during HMA therapy can predict response or resistance to treatment.
Following validation in the initial cohort, the model was externally validated in a cohort of 80 MDS patients treated with HMAs. The total of 494 patients had a median age of 72, and 71% were male. The Revised International Prognostic Scoring System (IPSS-R) scores for patients at the time of treatment were: 4% very low, 21% low, 24% intermediate, 21% high and 22% very high. The responses to treatment included: 11% complete remission (CR), 3% marrow CR, 3% partial remission (PR) and 29% hematologic improvement (HI).
The area under the curve (AUC) was used to evaluate the performance of the final model. A feature importance algorithm was used to identify the variables that most impacted the algorithm’s decision for a given patient.
When trained exclusively on serial CBC values (adding other clinical or molecular values did not improve the model’s performance), the model achieved an AUC of 0.82 in a cross-validated train/test schema and a similar AUC of 0.78 when it was applied to the second cohort.
Feature importance algorithms identified the following predictors of response: improvements in hemoglobin from baseline between days 21-30 of therapy, improvement in platelets between days 51-60, changes in monocyte percentage between days 41-50, and changes in mean corpuscular volume and red cell distribution width between days 31-60. The model also can create a personalized heatmap for a given patient that summarizes the variables that impacted their response or resistance to HMAs.
The MDS HMA treatment model has significant implications for clinical decision-making, offering oncologists a tool that can predict with high accuracy an individual patient’s response to treatment. “We could start patients on one HMA and, if it is not effective at three months, we can add a second drug or an investigational agent. Patients wouldn’t continue on a regimen that isn’t addressing their disease and would be spared unnecessary toxicity,” says Dr. Nazha.
Predictive models are an integral part of personalized medicine, enabling not only treatment designed for the individual patient but also “personalized prognosis.” Together, they can improve patient outcomes and minimize ineffective treatment.
The Center for Clinical Artificial Intelligence is working on additional models for hematologic and other cancers that will help realize this vision. “We are using novel, state-of-the-art artificial intelligence to address important clinical problems and building the future of medicine,” says Dr. Nazha.
Image: Confocal image showing the accumulation of myeloid hematopoietic cells throughout the mesenteric adipose tissue. Credit: National Institute of Allergy and Infectious Diseases, National Institutes of Health. Licensed. No changes were made.