Data are sourced from the AML17 [17] drug trial for patients younger than 60 years, which includes 3142 clinical records. Patient clinical records are combined with MRD (n=2587), NGS mutation profiles (n=3579), and a separate collection of NPM1 [6] and FLT3 specific mutation profiles (n=3142) [44]. A pseudonymization process converted patient trial IDs into dummy IDs before data access, ensuring compliance with participant privacy and data protection regulations. Access to the pseudonymized dataset was stored on the Cardiff University Research Data Store [45] with access restricted to authorized researchers. Clinical records contain measures at induction, including previous blood disorders, height, weight, the French-American-British 8-category AML classification [46], WHO and Eastern Cooperative Oncology Group performance status [47], cytogenetic, karyotype, ethnic background, and more. Metadata references to all used data are available in Multimedia Appendices 1-4.

After early diagnostic and comorbidity measurements during the induction stage, the proceeding 4 stages contain longitudinal records on periodic treatment, response, toxicity, and supportive care. The MRD subset includes quantitative polymerase chain reaction on peripheral blood and bone marrow samples, and multiparameter flow cytometry using the leukemia-associated immunophenotype and “different from normal” techniques. MRD measurements are longitudinal; the trial protocol involved readings at the end of each major trial stage investigated in this study. AML17 collected FLT3 and NPM1 mutation profiles with automated flow cytometry and manual bone marrow and peripheral blood cytology measurements at days 2‐3 from patient induction to the trial.

Access to data from the UK-NCRI AML17 clinical trial was provided by the Cardiff University Centre for Trials Research [48], which curates and governs the trial database.

All data were pseudonymized before being released to the research team. No direct patient identifiers were included in the dataset. All analyses were conducted on secure Cardiff University computing infrastructure. Ethical approval for the use of these data was granted by the Cardiff University School of Computer Science and Informatics Research Ethics Committee on February 28, 2025 (approval COMS/Ethics/2024/014). The research used secondary analysis of previously collected clinical trial data and did not involve direct contact with participants. Participants in the original AML17 trial provided written informed consent for their data to be used for research purposes. No compensation was provided to participants for this study, as it involved secondary analysis of previously collected trial data.

A diagrammatic summary of the model training pipeline, briefly describing the 4-phase process detailed before, is presented in Figure 1. Descriptions of CoxNet and RSF are detailed in Multimedia Appendix 6 and Multimedia Appendix 7. Code for relevant experiments is available in Multimedia Appendix 8.

Results are composed of, first, event (patient death within censoring window) status and timing distributions are measured, describing the spread of target events across cohorts and time for RSF and CoxNet models. Second, a Kaplan-Meier survival curve is shown for each cohort, including the full AML17 cohort, after cleaning for erroneous TTE indicators described in the Data Cleaning section. These curves provide a visual summary of the baseline survival patterns between cohorts, which the stage-specific survival models are tasked with capturing. Third, feature missingness correlations are visualized using heatmaps for data sources across AML17 cohorts, highlighting potential missingness mechanisms and motivating the usage of missing indicator variables for models. Fourth, Table 1 reports feature set sizes before and after reduction steps for RSF and CoxNet models, showing how reduction methods remove redundant features, focusing on the most informative and stable predictors quantitatively selected during phase 3 of the methodology. Fifth, c-index, dynamic AUC, and dynamic Brier quantify ranking accuracy, time-dependent discrimination, and overall predictive accuracy for each stage, providing a suite of metrics for cross-reference with similar work and reproducibility. Finally, feature importance is visualized using Venn diagrams of overlapping top-ranked features between stage-specific RSF and CoxNet models, and vertical bar charts illustrate the relative importance of the top 30 highest-ranking predictors in each model.

TTE (event=death status) of both models is the predictor variable of both models. Class imbalances influence the overall performances of a model [60]. There were no major class distribution imbalances for each of the 5 trial stages used for model training. No training sets had minority class percentages <41% and the average minority class percentage was approximately 44% (induction=42.8%, C1=46.9%, C2=48.9%, C3=41.3%, C4=42.9%). In literature, there is no definitive threshold at which imbalances are considered severe enough to affect the performance of ML models. A rule of thumb is that an imbalance is considered “moderate” when minority classes are 1%‐20% of the dataset [61]. Since this was not the case, the use of over- or undersampling techniques or synthetic data generation techniques such as the Synthetic Minority Oversampling Technique [62] was deemed not necessary.

Analysis of event time distributions in Figure 2 shows a disproportionate number of right-censored events for patient sets used for each of the 5 selected trial stages. This initially justified the usage of a c-index based on IPCW, which is specifically adapted for this situation [63] rather than the standard c-index, which is dependent on TTE distributions. However, it was found that differences between c-index and adapted IPCW c-index readings of CoxNet and RSF models were identical for stage models (eg, postinduction stage mean c-index=0.6760, mean IPCW c-index=0.6760). Full performance metric results across all stage ablation models are included in Multimedia Appendix 9.

Kaplan-Meier survival curves for each patient cohort for modeling show distinct survival patterns between trial stage cohorts, excluding the full cohort upon trial entry and the postinduction stage, as it began within 1‐3 days according to the AML17 protocol. This gives an indication of the distinguishable baseline survival patterns that vary between AML17 stage cohorts, which each model is tasked to predict. For visual clarity, the posttreatment stage 4 cohort was excluded from Figure 3, as it overlaps across many cohorts. Note that the posttreatment stage 4 was excluded from further modeling due to small sample size concerns (n=177).

Data missingness, nonexistent trial records, per trial stage cohort, and full AML17 cohort (post–data cleaning record exclusions) have been visualized using heatmaps for missingness correlation between features and frequency missingness plots for overall record missingness. Given the volume of total features, each figure has been further divided by data subset (eg, Clinical or NGS). The following matrix shows the missingness correlation between all clinical variables across the whole AML17 cohort used in this study. This provides justification for the usage of the missingness indicator features supplied as additional predictors for each model, which can be used to assess the informativeness of missingness and additional follow-up analysis. All additional figures have been included in Multimedia Appendix 10.

Across the entire AML17 cohort, the clinical subset has strong missingness correlations, suggesting a possible missing not at random mechanism for clinical data recorded at specific trial stages, indicated by the dark blue blocks of longitudinal features recorded at each trial stage in Figure 4. This is expected for any patients failing to proceed to a stage due to death or exclusion criteria.

When restricted to individual stage cohorts, the missingness correlation of clinical data shows a less pronounced block of strong positive relations, as this stage excludes patients who died beforehand (and thus have missing records) or who were otherwise no longer eligible based on trial protocols. Correlations that remain often are clinically explainable, for instance, records for comorbidity timings, such as those indicating nausea durations (“NauseastartC1” and “NauseastopC1”), are obviously both missing if the patient did not have such a symptom. However, to avoid possible assumptions, such variables, even if frequently missing across entire cohorts, were investigated for informativeness to model survival predictions by including an additional missing indicator variable for each; informativeness was then identified through stable feature importance analysis of final models illustrated in the Feature Importances section.

The highest performing models selected via c-index and feature stability selection defined in phase 3 (refer to Final Evaluation Phase in Methods section) were retrained for hyperparameter tuning under nested CV and re-evaluated according to c-index. Key statistics, such as training sample size, the ablation components, and feature set pre- and postreduction, are measured. Feature reduction was model-specific. RSF pruned features with average relative permutation importance. Nonzero feature coefficients determined by elastic net regularization remained in the CPHR model. Tables 1 and 2 show feature reductions, best performing data source ablations, training sample sizes, and average c-index with 95% CIs of each stage-specific model. Raw JSON metric files can be found in Multimedia Appendix 11.

RSF shows higher c-index readings across all stages. However, CoxNet had smaller CIs across all stages, suggesting it may be more generalizable, which requires further analysis using an external dataset. Given that the bootstrapping across each nested fold was set to 150, which was used as a compromise between precision and computational runtimes, it may be the case that a higher number of sample iterations could yield more precise and potentially smaller intervals in both models. All recorded evaluation metrics for final models trained with nested CV are available in Multimedia Appendix 11.

Using the full posttrial stage C1 cohort, the AML17 trial protocol risk assessment Cox linear regression model was compared with the RSF and CoxNet models at the same stage using a 0.9:0.1 train-test split and 1000 bootstrapped samples by c-index.

These results show the potential performance gain between trial risk-based models, such as the one used in AML17 for patient treatment stratification, and the RSF and CoxNet models used within this study. The reader is reminded that this specific protocol comparative analysis was exploratory and excluded a nested CV process, unlike those used to produce results highlighted in Table 2. The larger training sample set used for these specific models, while suggesting nonpessimistic performance with more realistic training set sizes, is at risk of overfitting against an external dataset and should be interpreted with caution. It should also be noted that CIs between all models overlap. Future work will include obtaining additional trial data from subsequent AML18 and AML19 trials as an external validation source.

Dynamic AUC was computed for each stage-specific model (RSF and CoxNet) from the beginning of the stage to the 5-year censoring point. Dynamic AUC quantifies a model’s discriminative ability at a given point in time t, representing how well the model distinguishes between patients who experience an event before t and those who remain event-free beyond t. AUC scores are bounded between values 0 and 1 inclusively, with higher values indicating better discrimination. A value of 0.5 corresponds to random chance, where a model cannot distinguish between patients. Values below 0.5 indicate a model makes predictions in the opposite ordering, effectively inverting the risk estimate. Note that an AUC of 0 indicates that the model perfectly ranks patients in the reverse order of risk; in such cases, inverting predicted risk scores would yield an AUC of 1, corresponding to perfect discrimination.

Following the ablation study, candidate models were evaluated by their mean c-index across all repeated, stratified folds of the CV process. At each trial stage, the highest performing RSF and CoxNet models were selected for a final, more robust nested CV with additional bootstrapping per fold for performance estimates. As observed, event time points differ across bootstrapping samples, a universal set of equally intervaled time points was first predefined. Each fold bootstrap sample AUC measurement was then interpolated to its nearest universal time point.

AUC shows an initial period of instability immediately after the stage’s baseline. This behavior was expected—early time points will have a smaller cumulative observed event count than the remaining time window of the trial stage. Additionally, risk distributions shift rapidly in these early time windows as patients transition into new treatments based on the existing protocol risk assessment and randomized allocation. As time progresses, all models tend to a more stabilized AUC score, shown by the plateau of the mean AUC and narrowing of CIs.

Later stages, (particularly poststage 3) exhibit visibly wider CIs and flatter trajectories. This reflects the considerably lower sample sizes of this stage (“Post-C3” n=761, Table 2) and higher censoring proportions (Figure 2) at this deeper trial stage. Consequently, this limits statistical power and increases uncertainty in time-dependent discrimination estimates. Therefore, dynamic AUC curves at poststage 3 (Figures 5G and 5H) should be interpreted with greater caution.

Dynamic Brier loss was computed for each stage-specific RSF and CoxNet model selected via the ablation study to evaluate prediction accuracy over time. At each time point t, the Brier score measures the mean square difference between the predicted survival probability at t and the observed event status. Score values are bounded between 0 and 1 inclusively, where 0 represents a perfectly accurate model and 1 a completely inaccurate model.

To enable consistent comparison across CV fold bootstrapped samples, dynamic Brier loss curves were evaluated over the same set of predefined, equally intervaled time points used for the dynamic AUC assessment.

Across all stages, dynamic Brier loss scores begin with very low loss values and narrow CIs. This behavior is expected—a smaller proportion of events has been observed at these baseline windows, most patients remain event-free, resulting in highly accurate short-term predictions at low variance. As time progresses throughout each stage model, the Brier loss score increases and the CIs widen. This reflects both the increasing difficulty of long-term survival prediction and the gradual decrease of the at-risk population contributing to the estimate (as the number of patients experiencing an event before t increases, decreasing the remaining subset of at-risk patients available from t).

Similarly, to dynamic AUC results, later stages (particularly poststage 3) show visibly wider CIs. This pattern arises from the reduced sample size available within the cohort (761 patients; Table 2) and higher censoring proportions (Figure 2), which limit the precision of time-dependent accuracy estimates and increase variance at later t evaluation points. Therefore, poststage 3 (Figure 6G and 6H) should be interpreted with caution.

Full resolution dynamic Brier loss and AUC plots are available in Multimedia Appendix 11.

The following figures show the feature importance of the best-performing tuned models at each select stage of the trial. Importance was ranked according to relative importance scores calculated from permutation importance for RSF and by coefficient values from CoxNet. For postinduction, post-C1, post-C2, and post-C3 models, RSF’s feature selection method included more features in the final model. Given the small difference in c-index and dynamic AUC performance between RSF and CoxNet during trial stage predictions, it is unlikely that the increased feature pool holds strong predictive candidates, suggesting that the increased RSF feature dimensionality from the feature selection process could pose a risk to overfitting and lack of generalizability.

Feature importance for each trial stage model has been ranked to highlight the strongest predictors. For figure readability below, only the top 30 highest-ranked features are included; the total number of features for all stage models exceeds this number, as illustrated in Table 1. While the principal goal of this study concerns the performance of time-sequential TTE prediction models through the trial, future works investigating and explaining generalizations of these models are necessary. Therefore, as a precursory step for future work, we have highlighted the most important features of the best models for each stage in Figure 7. Features that consistently mapped over all or the majority of (often excluding the early postinduction stage as longitudinal records did not exist at this time point) trial stage models include known AML prognosticators—age, white blood cell count, marrow blast values, cytogenetic risk groups (such as the core-binding factor t(8;21)/inv(16)), NPM1, and FLT3 markers [13,64,65]. There are also several consistently important predictors unreferenced as biological risk factors. Missing indicators (eg, blast marrow, FLT3 mutation type, and internal tandem duplication length value entries) suggest predictive informativeness of missing measurements, which may reflect selection bias, measurement practices from clinical sites, or the severity of a patient’s condition. Other predictors include administrative timing fields, such as marrow blast test dates or chemotherapy timings, which, while not strictly biological risk factors, suggest the importance of treatment timing or intensity. It is possible that more precise dosage levels and treatment timings could be recommended from such a system on a per-patient basis. Predictors measured as important but not referenced within the literature most likely reflect measurement practice bias or even data leakage, not necessarily causal disease biology, and warrant further analysis.

The primary performance metric, c-index, of the best models for each trial stage shows that RSF outperforms CoxNet throughout all major stages of the trial, with differences in performance becoming more apparent throughout successive trial stages. This suggests the optimal simulation of TTE outcome prediction would incorporate RSF or additional nonlinear model types that can capture nonlinear, complex relationships.

Brier loss of stage models appears similar in mean square error trends for all stage models except at stage 3, likely due to having the smallest sample size for training. In the instance of stage 3, the difference in accuracy is larger between RSF and its lower accuracy counterpart, CoxNet. However, it must be noted that the sample size for training of this model is the lowest, as final trial arm branches become increasingly fractional, using 761 samples. Therefore, overfitting is increasingly more likely at this stage, and conclusions made for the outperforming model should be made with caution.

Cumulative AUC shows that both models tend to perform similarly throughout time, being most unstable at the initial stages of the model’s available prediction window before censoring (which spans from the end of the respective stage to the 5-year censoring point since patient induction). Notable inflection points occur at similar early time frame windows, which eventually plateau into a stable time-dependent prediction. The degree of change during the initial window of each stage model before stabilization appears more drastic for both models, showing instances of under- and overperformance relative to their mean AUC score. This indicates that the initial periods of all trial stage models are the most sensitive zones in terms of predictive performance. While it cannot necessarily be concluded that the large performance differences before plateaus are solely a result of inadequate sample sizes, there are naturally fewer total event instances in the earliest sample periods with respect to the aggregate of events throughout the entire 5-year window. With low event counts at the earliest AUC time points, variability may be higher, resulting in an overestimate of performance as seen across dynamic AUC plots of all stage models. This would explain the noticeable fluctuations before stabilization of the mean AUC value for each stage model. This initial window would also be the most dynamic point in time with respect to patient treatment selection, where the trial protocol outline determined the treatment stratification of patients into one of multiple arms and so was highly influential on overall survival. This initial unstable period approximately coincides with the worst-case trial length for patients from induction to the end of stage 4, roughly 230 days [16]. Aside from effective sample sizes, it seems intuitive that the most sensitive prediction period of models trained on longitudinal data would be the initial time window after their measurements, as they more accurately reflect the real state of the patient, in the sense that there has been less time for longitudinal measurements to differ from their latest recorded value.

The ablation study stresses the importance of data held in the clinical dataset as well as FLT3 and NPM1 and longitudinal MRD records (metadata available in Multimedia Appendices 1-3), which were used for all the highest-ranked RSF models for each stage. The differences in ablated data sources will provide contextual pointers for future analysis of model features with traditional techniques (such as survival curves) or ML techniques (such as clustering), prioritizing features ranked most important by their respective model (Figure 1). Surprisingly, the NGS subset was only included in the postinduction stage RSF model, indicating that the strongest mutational markers came from the FLT3 and NPM1 dataset and karyotype information held in the clinical dataset, both of which remained consistently important across all stage CoxNet and RSF models, as seen in Table 2. NGS mutational biomarkers remained sparse per cohort and most likely effectively acted as noise to both models. This suggests further sensitivity analysis with a wider range of available features via composite NGS feature engineering. The preprocessing of NGS data in these experiments only used available gene mutation indicators to avoid catastrophic explosions in feature set dimensionality by the creation of composite variables. Some excluded features include tumor variant allele frequency and gene mutation base start and end locations, which are candidates for future analysis. A list of all available NGS features can be found in Multimedia Appendices 1-3.

Performance analysis of models constrained to data available up to trial stages approximately equates to performances in literature for a Cox proportional hazard model used for risk group stratification based on aggregate non–time point constrained data [26]. In the wider context of digital twins [43], the results in performance in this study, particularly at the earlier stage time points, suggest that it is possible to provide accurate simulations on survival risk based on models trained at iterative stages of treatment. A digital twin would simulate multiple AML patient outcomes that are relevant to patient well-being; this invites further study on other time-based outcome predictions using the generalized method of this study, such as for QoL or comorbidity prediction.

Approaching a digital twin system with a core prediction layer using ML models should have access to longitudinal data with a temporal resolution, optimal for AML-based predictions. Given longitudinal records, such as MRD, were used by the highest performing stage models (excluding the initial postinduction stage), future work should also involve sensitivity analyses on model prediction accuracy using more frequent, intrastage measurement. This is a practically challenging area of research as trial-based data, such as AML17, even with longitudinal records, is often restricted to stage-wise updates which last several weeks or more depending on chemotherapy regimen. Such a system also demands an automated, digitalized data collection scheme, which is not the norm in trial protocols, where many records are paper-based. Currently existing ML models trained on stage-wise model, such as those shown in these experiments, indicate the promising practicality of their usage in clinical environments when used as the core stratification method within a digital twin system. However, the surrounding architecture necessary to feed models with accurate patient data with high temporal resolution is lacking. The total man-hours to preprocess trial-specific records and validate models with comparable features across trial or real-world data sources for such a system would greatly compromise its applicability, unless a standardized data capture process which records data in an ML model–friendly manner (eg, with improved error handling – particularly for critical date-time entries, mandatory classification levels instead of clinician written note fields, stricter numerical field unit standardizations, a higher volume of QoL records, and higher resolution of longitudinal record entries) is developed for patients with AML. While RSF can model complex nonlinear relationships, they do not explicitly account for sequential dependencies; when higher resolution longitudinal data are available, researchers should evaluate the performance and feasibility of ML models designed to exploit temporal structures, such as recurrent or long short-term memory neural networks.

In the context of an AML digital twin, the presented RSF and CoxNet models here would act as the core prediction layer, updating patient-specific risk estimates whenever new measurements become available. As AML data are collected at discrete protocol-defined intervals (per trial treatment stages), this framework does not make use of real-time streamed data (which typically do not exist in trial-based datasets) but instead “real-time upon update” recalculation of risk as new clinical or MRD entries are recorded for a patient. Generalizability across clinical sites would be supported by a standardized preprocessing pipeline like what has been developed in this study. The monitoring of feature distribution drift (eg, via Population Stability Index) would allow for precise trigger points for model retraining when necessary. The planned external validation of RSF and CoxNet models as core digital twin predictive layers on AML18 and 19 datasets will further quantify cross-site robustness.

Model training and validation were computed on the Cardiff University “Hawk” high-performance computing cluster. Jobs were submitted via Simple Linux Utility for Resource Management, using 1 node, 1 task, and 14 CPU cores with the high-throughput partition. Processing time measurements show that RSF is much more intensive to validate than CoxNet, primarily due to the time complexity of the exhaustive feature importance method used—permutation importance. For example, when comparing runtimes of the largest cohort and ablation set—postinduction stage, with ablation components (clinical, MRD, NGS, FLT3, and NPM1)—averaged across the 3 repeated, k=5 split CV loop, RSF took on average 147 seconds per fold, while CoxNet on average took 1 second. This difference is made more apparent after the more robust internal nested CV process, where the same RSF and CoxNet models and ablations took 7.9 hours and 1.4 hours, respectively, for the entire validation process across all folds. A caveat arises, particularly with the usage of the RSF model in clinical applications with large sample sizes; while predictions from trained models are near instantaneous, the full training and validation times of such models may be costly for critical patients relying on fast treatment delivery. In practice, it may be necessary to use CoxNet (or a similar “fast” baseline model) for predictions and counterfactual simulations as a preliminary tool while more sensitive, albeit slower models, such as RSF, are trained. With respect to identifying when models should be retrained, a quantifiable approach to determine the threshold should be calculated. For example, the Population Stability Index can be used to measure if there is a significant drift in a feature between the reference (trained) sample set and the new dataset. This would minimize redundancy and processing constraints in retraining by otherwise using an arbitrary threshold for determining when a model should be retrained, in turn minimizing potential performance or generalizability loss.

Using a Cox proportional hazards model with ridge regression, Tazi et al [26] recorded an IPCW c-index of approximately 0.7 with a combined total of 26 clinical, demographic, FLT3 (ITD), and other molecular class features aggregated from UK-NCRI, such as AML11, 12, 14, 15, 16, and 17 trials. Models constrained to data at each trial stage in this study have comparable IPCW c-index scores of around 0.69 using RSF. The number of features retained in this study after model reduction processes is variable per stage, highlighting the significance of specific features at select time points for TTE predictions. Many features consistently shared across stages are noted in existing ELN risk classification, and those not referenced indicate the potential importance of timing and informativeness from both administrative and testing fields. The total number of retained features is higher for each stage model than in the study by Tazi et al [26], which used 26 features, as opposed to over 160 features across all RSF stage models. Future analysis will involve all features used at each stage in comparison with existing literature and guidelines for risk stratification, in addition to external validation to assess model generalizability.

RSF and CoxNet models outperform the AML17 protocol’s Cox model for risk stratification based on c-index (“Evaluation of AML17 Protocol Model” in Table 3, showing c-index readings of AML17 Protocol Cox Linear Regression=0.66, CoxNet=0.68, and RSF=0.70), suggesting both implementations can more accurately predict TTE. However, CIs overlap, suggesting larger datasets may be necessary to conclude an absolute improvement in generalized performance.

Since the completion of the AML17 trial, substantial progress has been made in patient therapy options through the identification of prognostic, predictive, and targetable molecular abnormalities [19]. While the analysis of models used in this study provides insight into survival prediction performance using longitudinally restricted data, future work would benefit from the inclusion of more recent trial datasets using newly approved first-line treatment options, such as Midostaurin and CPX-351. Additionally, AML17 patient-reported outcomes, including QoL, were excluded from model training due to concerns on overall sample size within trial stage time points. The future inclusion of datasets with larger pools of patient-reported outcomes data is of particular interest for the prediction of additional outcome responses that may reflect on the social and psychological health of patients at different stages of disease treatment and progression. Further research into additional outcome predictions can be integrated as part of a generalized AML digital twin that can inform patients and provide accurate recommendations to health care practitioners, particularly where survival risk between treatments is marginal, but there are clear differences in secondary predictions, such as patients’ QoL.

This study shows the practicality of 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. The performance of these sequential TTE models is intended to justify their use as part of a wider digital twin system simulating multiple TTE outcomes for patients with AML. The primary c-index metric shows comparable scores to the literature that uses similar models on aggregate sets of similar trial data. Consistent and stable important features relative to each stage-specific model are supported by ELN literature on AML classification, and additional nonreferenced predictors suggest the importance of stage-specific administrative and timing fields. Additional cumulative-dynamic AUC and Brier loss metrics have been provided. The most immediate future work includes feature analysis of the best models at each stage, further comparison with existing risk group stratification guidelines such as ELN, external validation with follow-up AML18 and 19 trial programs, the implementation of a minimally adapted pipeline for different outcome measurements, and the inclusion of patient self-assessed QoL form records alongside a broader collection of longitudinal MRD data, which have been recorded after trial treatment stages as detailed in the AML17 protocol and follow-up AML18 and 19 trials.