Predicting Symptom Presence Using Prior Health Indicators: Insights from the VISION Study

<strong><u>Background</u></strong> Accurate prediction of symptom presence following acute COVID-19 infection has important implications for clinical monitoring and patient management. The Vital Status and Outcomes of COVID-19 (VISION) Study provides longitudinal self-reported symptom data that enables evaluation of machine learning approaches for predicting recovery trajectories. This study aims to assess the performance of classical machine learning and deep learning models in predicting symptom status 14 days following recent self-reported SARS-CoV-2 infection or a return of COVID-19 symptoms within 7 days of initial improvement and completion of oral COVID-19 treatment through the use of baseline and longitudinal participant-reported health information. &nbsp; <strong><u>Methods</u></strong> Participants from Arm 1 of the VISION study with completed baseline surveys and acute symptom diaries were included. Three modeling approaches were implemented: (1) baseline and Day 1 features only, (2) baseline features with two longitudinal health indicators across varying observation windows (Days 1-7 through Days 1-13), and (3) the same feature set applied within symptom subgroups defined by symptoms reported on Day 1 of diary entry. Elastic Net, Random Forest, XGBoost, and multi-layer perceptron (MLP) models were trained using a wide-format representation of longitudinal data, while a long short-term memory (LSTM) model was trained using sequential input data. Final model performance was evaluated using accuracy and area under the receiver operating characteristic curve (AUC) on the held-out test data. <strong>&nbsp;</strong> <strong><u>Results</u></strong> Across all approaches, inclusion of longitudinal symptom data improved predictive performance relative to models including only Day 1 information. In Approach 1, Elastic Net achieved the highest test AUC (0.678) with limited overall performance. In Approach 2, models trained on longer observation windows (e.g., Days 1-13) demonstrated improved performance, with XGBoost achieving the highest test accuracy (0.851) and Elastic Net achieving highest AUC (0.903). Performance across classical machine learning models remained generally similar, while LSTM did not outperform the classical models across observation windows. Feature importance from XGBoost models indicated that return to pre-COVID health status was the most consistently influential predictor, particularly when measured at later days within the observation window closer to the Day 14 prediction target. In Approach 3, subgroup analyses showed improved performance in larger symptom subgroups, with highest test AUC observed for Random Forest (0.921) and greater variability in smaller subgroups. MLP consistently demonstrated lower performance relative to other models. <strong>&nbsp;</strong> <strong><u>Conclusions</u></strong> Incorporating more recent symptom information improves prediction of Day 14 symptom status. However, model performance is strongly influenced by data structure and sample size. Classical machine learning methods, particularly Elastic Net and tree-based models, performed comparably or better than deep learning approaches in this setting. These findings may suggest the usefulness of interpretable predictive modeling methods that may be preferable for predicting symptom trajectories in moderately sized and self-reported longitudinal health datasets.<strong><br /></strong>