Sequential model for predicting patient adherence in subcutaneous immunotherapy for allergic rhinitis

Area of Science:

• Allergy and Immunology
• Artificial Intelligence
• Health Informatics

Background:

• Subcutaneous Immunotherapy (SCIT) offers long-lasting treatment for allergic rhinitis (AR).
• Patient adherence is critical for maximizing the benefits of allergen immunotherapy (AIT).
• Predicting non-adherence and symptom scores is essential for effective AIT management.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting SCIT non-adherence and local symptom scores in AR patients.
• To compare the performance of sequential latent-variable models (SLVM) based on Stochastic Latent Actor-Critic (SLAC) and Long Short-Term Memory (LSTM) networks.

Main Methods:

• Utilized SLAC and LSTM, recurrent neural networks designed for time-series data, to model patient adherence dynamics.
• Evaluated models based on their ability to predict adherence and local symptom scores over a 3-year SCIT period.
• Excluded biased samples from the initial time step for robust analysis.

Main Results:

• LSTM models achieved predictive adherence accuracy ranging from 66% to 84%, outperforming SLAC models (60% to 72%).
• Root Mean Square Errors (RMSE) for LSTM were between 1.09 and 1.77, while for SLAC they ranged from 0.93 to 2.22.
• Both models demonstrated significantly lower RMSE than random prediction error (4.55).

Conclusions:

• Sequential models offer a promising approach for the long-term management of SCIT in AR.
• LSTM demonstrated superior performance in predicting patient adherence, while SLAC excelled in symptom score prediction.
• The SLAC model's flexibility provides a novel method for managing long-term AIT.