Introduction
Subcutaneous Immunotherapy (SCIT) represents one of the few disease-modifying treatments for Allergic Rhinitis (AR), aimed at inducing long-term immune tolerance. Unlike symptomatic treatments, SCIT requires a significant commitment—typically three to five years of regular injections. This "marathon" nature of treatment often leads to poor adherence, with dropout rates reported between 25% and 90% in various clinical settings.
Predicting which patients are likely to discontinue treatment is crucial for clinical intervention. While traditional predictive models often treat adherence as a simple binary outcome (completion vs. dropout), they fail to capture the dynamic nature of a multi-year treatment plan. This study explores the application of sequential machine learning models to provide step-by-step predictions of both patient adherence and symptom progression, allowing for timely interventions.
Methodology
The researchers developed and compared two distinct sequential architectures to model the patient journey over a 36-month period:
- Long Short-Term Memory (LSTM): A recurrent neural network designed for time-series data, predicting next-step symptom scores and adherence status simultaneously.
- Sequential Latent-Variable Model (SLVM): Based on the Stochastic Latent Actor-Critic (SLAC) framework. This model treats adherence as an "action" within a "state-action" pair, allowing it to simulate how different adherence behaviors influence future health states.
The study utilized a retrospective dataset of 205 AR patients from the First People’s Hospital of Foshan, China. Data points included 14 baseline variables (e.g., age, gender, distance to clinic, IgE levels) and 11 longitudinal variables (e.g., nasal/ocular symptom scores and medication scores). Observations were recorded at specific intervals: 0, 4, 12, 18, 24, and 36 months.
Results and Discussion
The evaluation revealed a specialization trade-off between the two models. The LSTM proved more robust for binary adherence prediction, achieving an accuracy range of 66% to 84%. In contrast, the SLAC-based model excelled at predicting numerical symptom severity (VAS scores), yielding a lower Root Mean Square Error (RMSE) of 0.93 to 2.22, compared to the LSTM's 1.09 to 1.77.
A significant advantage of the SLAC model is its capability as a clinical simulator. By modeling adherence as an action, clinicians can input hypothetical scenarios—such as a patient missing several months of treatment—to visualize predicted symptom regressions. This "what-if" capability serves as a powerful tool for patient education and shared decision-making, potentially motivating patients to stay the course.
Interpretability analysis using the Captum library highlighted that "distance to clinic" was the most significant predictor of non-adherence. This suggests that logistical and geographical barriers are often more influential than biological factors in determining long-term treatment success. Furthermore, the data identified the one-year mark as a critical "danger zone" for dropouts, a trend exacerbated by the COVID-19 pandemic, which accounted for 25% of first-year dropouts in this cohort.
Conclusion
This research demonstrates the potential of sequential latent-variable models in managing chronic conditions like Allergic Rhinitis. While LSTMs are effective for identifying at-risk patients, SLVMs offer deeper insights into symptom progression and the consequences of non-adherence. Future work should focus on expanding the cohort size, standardizing observation intervals, and integrating these predictive tools into digital health platforms to provide real-time clinical support and improve long-term patient outcomes.
Summary updated