Table 3.

Summary of challenges in applying ML techniques in HSCT.

Challenges	Reasons	Potential Solution
Limited Data Capture	■Complex HSCT procedure with numerous post-transplant complications ■Lack of continuous and real-time capture of various data streams involved ■Mix of automated and manual data capture	■Utilize wearable sensor devices or leverage mHealth platforms for robust data collection
Data Quality Issues	■Lot of missingness and inconsistencies due to complex data collection procedures ■Loss of important variables lead to loss of relevant information	■Developing autonomous, adaptive, and online preprocessing algorithms that can automatically capture the data quality issues and resolve them by employing appropriate techniques in real-time
High Dimensional Data	■Large number of clinical and/or genomic variables associated with the HSCT outcome	■Developing novel streaming dimension reduction techniques for efficient processing of large number of features associated with the HSCT outcome
Data Privacy Issues	■Large amount of sensitive patient data is required in building predictive models due to numerous factors involved ■Combining multiple data streams from disperse data stores leads to potential data privacy issues	■Developing appropriate privacy measures, such as data anonymization techniques to ensure complete privacy of patients’ data ■Using technique such as “federated learning” [47] that trains a shared global model via a centralized aggregation server, while keeping sensitive data in local institutions of their origin ■Enabling some form of privacy access control to different data streams that can ensure that only those with proper authorization can access a patient’s data streams
Obsolete Predictive Models	■Dynamic evolution of disease states in patients undergoing HSCT	■Developing adaptive ML techniques having capability of detecting data changes over time and adapting accordingly
Diverse Data Types	■Captured data are of different modalities and sampled at different rates	■Multi-modal data integration techniques using deep learning has to be developed for effective integration
Data Integration issues	■Most of the captured data are typically dispersed among various data stores (e.g., cloud storage, EHR, individually-managed databases)	■Using mHealth platforms could be a potential solution.