Deep Learning Models for Predicting Overall Survival of Acute Myeloid Leukemia Using Short-Term Longitudinal Blood Measurements and the Omop Common Data Model

Background

Accurate prediction of outcome of individual acute myeloid leukemia (AML) patients is crucial for guiding value-based treatment decisions and improving survival. However, existing prognostic models are based on static, retrospective data, mostly from clinical trials, which do not fully capture the current real-world population and treatment practice of AML.

Short term follow-up complete blood count (CBC) measurements are routinely taken, but current models do not account these longitudinal data and their prognostic value is unknown. Moreover, the lack of standardized and interoperable data across different health care systems hampers the development and validation of scalable, predictive models. These types of close-to-realtime models can be continuously and even semi-automatically updated and integrated into clinical decision systems.

Deep learning is a promising approach for building complex and accurate predictive models from rich and longitudinal real-world data (RWD), particularly when presented in a standardized common data model (OMOP CDM), which is being universally adopted (e.g. ASH Data Hub, The All of Us Research Program, DARWIN EU).

Aims

1. Test the feasibility of using OMOP RWD, including longitudinal, short-term follow up data, to build predictive models for precision medicine.
2. Develop and implement a modeling framework.
3. Provide a proof of concept by building a robust AML model to predict overall survival based on baseline and short-term follow up CBC measurements.

Methods

We used OMOP RWD from Helsinki University Hospital for deep learning models for predicting overall survival of AML patients. All patients diagnosed with AML between 2000 and 2022 and had at least three CBC measurements (incl leukocyte differential and LDH) within 21 days after diagnosis. Missing values were imputed with -1, allowing the model to learn that -1 means missing. Data were split into training, test, and validation sets. We used a two-layer fully connected artificial neural network (ANN) based on PyTorch as the base model, connected to a survival model called DeepHit as the output layer. DeepHit is a deep learning approach that can handle competing risks and provide individualized survival predictions. Dropout layer with a dropout rate of 0.4 were used to avoid the risk of overfitting. We trained the model using cross validation on the training split, and validated it on the independent test set. Kaplan Meier plots were created by splitting the cohort into three strata (upper, intermediate, and lower quantile) based on the model predicted patient-specific risk.

Results

After applying the inclusion criterion of having at least three CBC measurements within 21 days after diagnosis, 614 patients remained for the analysis. The data were split into training, validation, and test sets, of 393, 98, and 123 patients, accordingly. The model was trained on the training set using cross validation, and the best performing model was selected based on the validation set. The model converged quickly, requiring less than 100 epochs to reach the optimal performance. We experimented with different architectures of the ANN, varying the number of hidden layers and nodes. We found that the model performance was not sensitive to the architecture, and even a simple two-layer network with 16 nodes in each layer achieved good results. The final model was able to stratify the patients into three risk groups with markedly different survival probabilities. 2-year survival for the model-predicted low versus high-risk patient groups was 71% vs 10% in the training and 52% vs 11 % in the independent test cohort. Note that the model can also provide individualized survival probability curves for each patient and the uncertainty of the predictions over time.

Conclusion

Building robust predictive models based on longitudinal, close to real-time OMOP RWD is feasible and has great potential. Using dynamic, short term, universally available follow up data only – CBC measurements up to 21 days post diagnosis - holds valuable information for predicting long-term prognosis, including overall survival. These models will be further refined using more traditional baseline features (cytogenetics, mutation profiles, clinical characteristics) and iteratively matured in a multicenter network setting using advanced modeling and swarm learning, a form of distributed machine learning tailored towards medicine.