OBJECTIVES :

Acute Myeloid Leukemia is an aggressive malignancy with poor prognosis despite significant efforts to identify new therapies. Adverse events can account for 24-47% of clinical trial discontinuation¹. In this study we leverage real-world data and machine learning approaches to identify patient-level predictors of common adverse events in AML patients. We then validate the model’s accuracy in a clinical trial population to evaluate its potential utility for improving clinical trial conduct and helping patients avoid adverse events.

METHODS

De-identified Oncology Electronic Medical Record (EMR) data was used from the Guardian Research Network™ (GRN) of integrated delivery systems from Jan 1990 – July 2018. Pooled trial patients with relapsed/refractory AML were from Medidata Enterprise Data Store from March 2008 - Nov 2017³. All data sets were converted into the OMOP Common Data Model, v5. Analyses were conducted in SHYFT Quantum v7.1.1 and Python 3.6. All subjects had at least 180 days pre- and post-index activity, indexed at initiation of most recent chemotherapy regimen. Rates were calculated for adverse events commonly associated with treatment discontinuation – neutropenia, thrombocytopenia, anemia and pneumonia². Features included, but were not limited to, age, gender, current and prior chemotherapy, prior stem cell transplant, and support medications. Predictors were assessed using multiple machine learning algorithms (Logistic Regression, Gradient Boosted Classification, Random Forest, and XGBoost Classification). A 3:1 training/test split and cross validation scoring with a K-fold of 6 was employed. Model performance was measured by AUC ROC scores, PPV and NPV.

RESULTS

For EMR-derived models, AUC scores were 0.60-0.75 for top performing models. Testing in clinical trial data had somewhat lower performance, with AUC’s .56-.61. Best performance was seen in models for neutropenia.

CONCLUSIONS

Real-world data and machine learning can assess risk of potentially treatment-limiting adverse events, with potential applications for reducing clinical trial discontinuation and managing patient treatment burden