Predicting Platelet Demand

Challenge

Algorithms to improve the user experience

As experts in data-driven solutions, we often encounter complex logistical puzzles. One such challenge, deeply impactful and critically important, lies within the intricate world of healthcare supply chains. Imagine a scenario where a product with an extremely short shelf life requires meticulous planning and distribution to meet the ever-changing needs of patients. This was the challenge presented to us when we began looking into the complexities of blood platelet management. Organizations responsible for producing and supplying these crucial components for hospitals face a daunting task: reliably meeting the demand while simultaneously minimizing the waste caused by product expiration. The pressure to ensure a high delivery rate is significant; shortages can impact patient care, while overproduction results in financial losses and ethical concerns surrounding wasted donations.

The short shelf life of platelets, combined with the variable demand, creates a complex logistical challenge. Current methodologies at our client, frequently rooted in historical averages and static calculations, often fall short. The fluctuating nature of demand, influenced by factors such as holidays, seasonal trends, and unforeseen spikes in medical needs, makes accurate forecasting a crucial necessity. The inability to make last-minute production adjustments, given the 1-2 day lead time in the supply chain, only adds to the complexity. We recognized the need for a more sophisticated approach – one that moves beyond reactive responses to a proactive stance, leveraging the power of data to accurately predict demand. Our challenge was clear: develop a solution that could transform the management of blood platelets through precise and dynamic forecasting. This would not only optimize production and reduce waste but, most importantly, contribute to improved patient outcomes.

Approach

Understanding the Landscape: Initial Data Insights

Our approach to optimizing blood platelet management began with a deep dive into the available data, focusing on factors impacting demand such as holidays and seasonal trends. The data showed a clear seasonality in demand, however, it wasn't defined by extreme swings, and we noted that the COVID-19 pandemic required us to use data from 2022 onwards for training the model. We also determined that individual hospital models would not be viable due to data scarcity and sporadic ordering patterns; therefore a more holistic approach was needed. To establish a baseline, we analyzed our client's current forecasting method, which relies on the historical average demand per weekday. This served as the starting point against which we would measure the performance of our new, data-driven solution.

Model Experiments and Selection

Having established a solid baseline with the current forecasting method, we embarked on a series of model experiments, using a rigorous walking-forward cross-validation technique to account for the time-series nature of the data. Every experiment was meticulously logged using MLflow, ensuring full transparency and reproducibility. Our initial exploration included statistical models like (S)ARIMA and Prophet, commonly used for time-series forecasting. We considered ARIMA, but the complex non-linear patterns in the data meant it was not an ideal choice. Similarly, Prophet didn't capture the influence of holidays as effectively as LightGBM. We then moved on to decision tree-based approaches, such as Random Forest and LightGBM, exploring their capabilities to capture non-linear relationships in the data. Finally, we ventured into neural network architectures, including the Temporal Fusion Transformer (TFT), known for its ability to handle complex temporal patterns.

After a comparative analysis, the LightGBM model demonstrated superior performance, outperforming the client's existing forecasting method by approximately 27.5%, expressed in terms of the Root Mean Squared Error. LightGBM demonstrated consistently robust performance, especially when processing larger datasets, and showed improved accuracy as our training data grew over time. Unlike ARIMA and Prophet, LightGBM can also effectively handle more complex non-linear patterns present in the data by incorporating important covariates like holiday and seasonality-related features. LightGBM especially excelled over Random Forest, when dealing with longer-term predictions, thanks to its stronger generalization capabilities and reduced tendency for overfitting. Finally, LightGBM proved to be more robust against outliers, easier to tune, and had less demanding hardware requirements compared to the more computationally expensive TFT.

Understanding the model and End-to-End Implementation

Our LightGBM model's accuracy was enhanced by several key features. Historical demand, particularly the demand from one and seven days prior, proved most impactful for capturing both short-term fluctuations and weekly patterns. Seasonality also played a crucial role, with the day of the week and month of the year used to learn annual trends. Additionally, holiday-related features were important, including the distance to the previous and next holiday, and the categorization of each holiday as occurring at the start, middle, or end of the week, all contributing to better predictions of demand spikes and dips. The model's predictive power was further strengthened by incorporating past and known future events like public holidays. By using these future events, the model can refine its forecasts and enable proactive production planning.

To operationalize the solution, we implemented an end-to-end pipeline within Azure Databricks. This pipeline generates bi-weekly forecasts for the next 30 days, ensuring seamless integration with the client's existing systems for reporting. This automated process ensures a seamless flow of information, allowing the client's team to access the forecasts via their established reports and to adjust their stock accordingly. This integration is crucial for the practical application of our data-driven forecasting system, as it allows the client to consistently plan their production while minimizing the need to adjust their standard processes and systems.

Furthermore, we have included several monitoring and evaluation capabilities to guarantee the ongoing accuracy and dependability of our solution. We have implemented master notebooks using Azure Databricks, that allow us to run prediction for any specified period and evaluate how it aligns with actual usage for backtesting. The monitoring tools track the model’s performance and detect potential data or performance drift. Dynamic visualizations also allow for a deeper analysis of the model's performance and patterns. Moreover, we use NannyML to monitor for model drift, proactively detecting any shifts in data or decreases in performance, ensuring our model consistently delivers dependable and precise forecasts.

Impact

Our machine learning solution has significantly improved our client's blood platelet management, offering a substantial upgrade from their previous methods. By providing more accurate demand forecasts, we've achieved tangible operational and patient-related benefits. The improved forecasting accuracy – approximately 27.5% better than the baseline – translates to valuable enhancements across the supply chain.

Financial Savings & Efficiency: A more refined understanding of demand, especially around holidays, has a positive impact on production costs. While a 27.5% increase in forecast accuracy doesn't directly equate to a 27.5% reduction in waste (as other factors influence the overall process), even small improvements can yield considerable savings. Considering the cost of platelets—over 500 euros per unit—and their limited shelf life of 3-5 days, the previous system experienced the loss of more than 3000 platelets annually due to expiration. By reducing overproduction with more accurate forecasting, our model has the potential to realize significant cost savings for our client annually, resulting in a strong return on investment.

Improved Supply & Patient Care: One of the most critical results of our solution is its contribution to more consistent platelet supply. Shortages can lead to delays in necessary transfusions for patients undergoing surgery, cancer treatment, or managing chronic conditions. Our enhanced forecasting ensures a more reliable supply of platelets, which supports consistent patient care and avoids potential disruptions in treatment. The focus is on optimizing production to improve patient care through efficient resource allocation.

Supporting Blood Donation Efforts: Our solution also helps ensure that blood donations are used efficiently. By optimizing the usage of donated platelets, we aim to minimize waste and ensure that these contributions have maximum impact. This demonstrates a commitment to the valuable role of blood donors and reinforces the significance of community involvement in the healthcare supply chain.

Overall, this project demonstrates the potential of data-driven solutions to improve healthcare operations. Our ability to enhance resource allocation, improve supply reliability, and reduce waste has positive implications for both efficiency and patient well-being, while also providing business advantages for our client.