Deep learning for platelet transfusion

In this issue of Blood, Engelke et al¹ introduced a deep learning approach, specifically long short-term memory (LSTM) models, to predict the need for a platelet transfusion within the next 24 hours based on static and time-dependent patient data. Platelets have a short shelf-life, which makes logistics challenging to ensure their availability when and where needed. As a result, wastage is often high. Given that platelets are donated by volunteers and are costly to the community, there is a need to make better use of the available platelet supply. The models in the study by Engelke et al explore the potential of deep learning in predicting patient platelet transfusions.

Using the expanding window method, the models in Engelke et al considered the most recent 30 days of medical events, such as platelet count measurements and historical platelet transfusions. This approach allows for learning from both short- and long-term data sequences but has its challenges. For patients with prolonged stays, only their latest data may influence model predictions. Patients with short stays, due to death, transfer, or discharge, did not contribute a full 30-day data sequence to the modeling process. To maintain a uniform input length for all patients, sequences of <30 days were padded. This might lead to less accurate predictions for both long- and short-stay patients.² 

Engelke et al used an 80-20 data split: 80% for model training and 20% for testing. A subset of the training data was reserved for model validation. The input predictors for the model development were classified into 3 categories: (1) 7 time-dependent variables (eg, indicators for previous platelet count measurement and platelet transfusion) over 60 half-day intervals, (2) 2 location variables (ie, patient’s last department and ward), and (3) 16 patient-level characteristics (eg, age, procedures, and medications) within the past 30 days.

At the core of the model architecture, the dense layer processes all the input data. Then, using a function called “sigmoid” (a standard choice in machine learning), it generates a prediction score between 0 and 1. The closer the score is to 1, the higher the likelihood of a platelet transfusion being needed in the next 24 hours. The authors also applied several techniques to enhance the model performance and efficiency.

For model interpretation, they used the precision-recall curve, which was calculated across various thresholds within the data set, to determine a binary label for the outcome decision (ie, whether a platelet transfusion is required within the next 24 hours: yes or no). They also presented the relative feature contributions of the speciality-specific models and compared the performance with non–deep learning models. Although their approach provided valuable insights, there remains an opportunity to fully explore whether the strengths of the deep learning (LSTM) model have been leveraged: are there complex data patterns (nonlinearities) and/or delayed effects over time (long-lag dependencies) in the data?³ If these complexities are not present, simpler models may be more appropriate.

The authors used FHIR-PYrate⁴ for data extraction in a standardized format. In theory, once developed, the model should be applicable across a range of health care environments, from local hospitals to specialized tertiary centers. However, although the model design hints at broad adaptability, its real-world efficacy in diverse settings has yet to be proven. By building the models to adapt to local data, the authors assume that institutions possess profound expertise in both their data and deep learning, and are prepared to reproduce and fine-tune the model as needed. This might present substantial implementation hurdles in numerous health care facilities.⁵ These implementation hurdles may well be at a similar level to the operational research and simulation approaches mentioned earlier in Engelke et al, albeit with a different skill set.

The authors’ decision to develop specialty-specific models, particularly for high transfusion demand departments, such as hematology, oncology, and cardiothoracic surgery, underscores a thoughtful approach. However, with the possibility of patient transfers within health care facilities, one must consider whether this model's compartmentalization adequately captures the complexity of interdepartmental interactions and the subsequent implications for platelet demand.⁶ The notable performance disparity among departments, with an area under the precision-recall curve of 0.84 for hematology-oncology compared with a mere 0.41 for cardiothoracic surgery, accentuates these concerns. Furthermore, the unexpected finding that platelet count was not among the top predictors is both surprising and alarming, which reflects the well-known lack of transparency inherent in deep learning models.⁷ Although the authors suggested that the complexity of deep learning reflects the multifaceted nature of platelet transfusion prediction, clinicians may find it challenging to reconcile with a model that appears to underemphasize the significance of direct clinical markers for platelet transfusion.⁸ Hence, although the study presents a modern analytical approach to platelet demand management, the model's efficacy across varied health care settings requires rigorous validation.

The authors suggested that their models can enhance platelet management in several ways: identifying high-risk patients, improving inventory management, and integrating with existing platforms, such as GE HealthCare Command Center. Moreover, by combining their models with conventional forecasting techniques, there is potential for more agile responses to demand shifts, smoother platelet transfer processes, and optimized resource allocation between hospitals during emergencies. However, the question of whether the models in Engelke et al can be directly transitioned for broader platelet demand forecasting or inventory management requires further exploration. It is essential to recognize that these research areas demand distinct modeling strategies and expertise that differ from the patient-level predictions considered by the authors. Although predicting individual transfusion requirements holds merit, it does not equate directly to the broader aims of managing platelet inventories and resource allocation.⁹ Integrating both methods could open up new avenues.¹⁰ Yet, it is important for readers and policy makers to understand and appreciate the distinctions between them.

In conclusion, Engelke et al offer promising advancements in platelet transfusion predictions using deep learning. The models hold potential in enhancing platelet management and adapting to diverse health care settings. As we venture into its practical implementation and explore its accuracy across specialties, taking steps for rigorous validation and gaining a deeper understanding of its nuances will further elevate its innovative contributions.

Conflict-of-interest disclosure: The authors declare no competing financial interests.