JMIR Publications

ABSTRACT

Background:

Risk group stratification based on AML patient survival prediction is complex. Despite common risk group categorisation guidelines, overall prognosis remains poor. Machine learning (ML) techniques have been shown to provide more accurate risk group stratification than conventional approaches using trial data. However, many time-to-event models do not utilize training sets constrained to specific time windows, instead using aggregations of trial data.

Objective:

Evaluate the performance of 1) Random Survival Forest (RSF) and 2) Cox Proportional Hazard Regression (CPHR) with Elastic Net regularisation (CoxNet) for survival prediction of Acute Myeloid Leukaemia patients at discreet time points during the AML17 randomised controlled trial dataset.

Methods:

For each stage in the AML17 trial, patients were split into a 4:1 train-test split with separate models trained for each exhaustive k-choice combination of available AML17 data subsets. Data combinations for each model were further constrained according to the respective trial stage to avoid data leakage. Preliminary Pearson’s correlation methods were used to remove directly correlating features with the time-to-event prediction (time-to-death/5-year censoring point). Permutation importance and Elastic Net regularisation were used to reduce the feature set of RSF and CPHR models respectively. Average c-index based on inverse probability of censoring weights was the primary metric for baseline, intermediate and hyper-parameter tuned models using a 5-fold, randomly shuffled, cross validation approach. Intermediate models in optimisation method were evaluated by average concordance index (c-index) based on inverse probability of censoring weights (IPCW) across each five folds of the training set within a cross-validation method. Final models were evaluated using the held-out test set.

Results:

IPCW c-index ranked the best models for data constricted up to the end of induction (RSF: 0.69), stages 1 (CoxNet: 0.67), 2 (CoxNet: 0.65), 3 (CoxNet: 0.65), and 4 (RSF: 0.6463) of the trial.

Conclusions:

This study details the high prediction accuracy for time-to-survival-event predictions when training sets of CoxNet and RSF models which are sequentially constricted to data measured up to the end of respective AML17 trial stages. Performance of these sequential time-to-event models intend to justify their use as part of a wider digital twin system simulating multiple time-to-event outcomes for AML patients.