JMIR Publications

ABSTRACT

Background:

Treatment of acute myeloid leukemia (AML) for older or unfit patients typically involves treatment with venetoclax plus azacitidine (ven/aza). Toxicity and treatment responses are highly variable following treatment initiation and clinical decision making continually evolves in response to these as treatment progresses. To improve patient clinical decision support following treatment initiation, predictive models based on evolving and dynamic toxicities, disease responses and other features should be developed.

Objective:

We aimed to generate machine learning based predictive models including individual patient predictors of long-term outcomes based on early clinical events occurring after initiation of venetoclax ven/aza in patients with AML.

Methods:

We quantified the association between toxicities, hospital events such as ICU transfer and transfusions, and short-term disease responses after ven/aza with longer term responses and compared these to a control group of younger fitter patients treated with aggressive chemotherapy (7+3). Univariate and multiple penalized survival regressions were performed and key features for both treatments that were associated with long term survival were identified. From these, a series of machine learning (ML) based predictive models for individual patients’ mortality were developed for both arms. The treatment-specific optimum models were chosen with respect to discrimination ability of survival models using cross-validation. Uncertainty in parameter estimation and subject-specific, survival prediction over time were assessed by bootstrap-based approaches.

Results:

Ven/aza was associated with fewer and different toxicities than 7+3. Best responses occurred at later time points following ven/aza relative to 7+3. Short term and best responses had different associations between ven/aza and 7+3 and long-term outcomes including OS. For both treatments, ML based predictors of long-term outcomes demonstrated survival AUCs of >0.70 with certain ML models.

Conclusions:

Our results demonstrate that a variety of toxicities, clinical events and responses evolve over time and at different kinetics and with different long-term implications following initiation of treatment with ven/aza compared with 7+3 type treatments ML based predictive models were feasible as a clinical decision support strategy for both forms of AML treatment. If validated with larger and more diverse data sets, these findings may provide valuable clinical decision support for newly diagnosed AML patients based on clinical features that evolve following treatment imitation.