(115) Nirmatrelvir and Ritonavir (Paxlovid) Prognostication: Unveiling Utilization Trends with Neural Network Time Series Forecasting

Background: In response to the ongoing COVID-19 pandemic, nirmatrelvir and ritonavir has emerged as an antiviral treatment option for eligible individuals infected with COVID-19.

The Department of Veterans Affairs (VA) Pharmacy Benefits Management (PBM) Formulary Management office, leveraged a forecasting model to provide valuable insights into anticipated utilization patterns of Paxlovid to aid decision makers with resource allocation and distribution to ensure uninterrupted patient access.

Objectives: To develop a time-series forecasting model that can accurately predict the utilization patterns of Paxlovid within VA over a one-year period to provide actionable insights into resource distribution.

Methods: The historic dataset was prepared from a Microsoft SQL Server database containing registry data collected by PBM.

Tools used for data preprocessing, model development, and statistical analysis include: Python, NumPy, Pandas, Sci-kit Learn, TensorFlow/Keras, Plotly, and Matplotlib.

A custom hybrid Convolutional and Recurrent Neural Network (CNN/RNN) sequential model consisting of a Conv1D layer followed by multiple long short-term memory (LSTM) layers was chosen based on accuracy and flexibility after evaluating five deep learning models using mean square error (MSE) test.

The model is designed to provide a robust framework for time series forecasting and is optimized using the Adam optimizer with a specified learning rate.

Model performance during the training process was monitored using both the training and validation datasets over 100 epochs with results used to fine-tune model parameters.

Results: The custom model made predictions for 1 year beginning 10/28/2023. Predicted utilization was 11,110 Paxlovid courses, with an actual utilization of 11,725 resulting in a -5.25% variance over the past 100 days. When excluding outlier data resulting from a 2-week holiday surge caused by external factors from 12/24/2023 to 1/6/2024, aggregated utilization was 9,196 with 9,137 courses predicted with a -0.63% variance over the same time window.

Conclusions: The developed custom hybrid prediction model is in close agreement with actual courses utilized and demonstrates a comprehensive approach to forecasting future doses based on historical patterns. The custom hybrid RNN/CNN design allows the model to capture both local and global temporal dependencies, making it suitable for the intricate dynamics of time series data.