Poster M4 - Machine Learning and Electronic Health Records Predict tDCS Treatment Outcomes of Anxiety in Older Adults

Introduction: Anxiety disorders significantly impact public health and may exacerbate conditions like dementia. Transcranial direct current stimulation (tDCS), a non-invasive brain modulation technique, offers potential relief but exhibits variable efficacy across individuals. The variability in tDCS outcomes underlines the need for better predictive tools to tailor treatments to individual responses. This study utilizes Lasso regression to enhance predictive models analyzing data from 333 participants, aiming to identify when clinically significant improvements in anxiety might occur. This predictive approach aims to enhance the accuracy of our models and establish a foundation for precision dosing. By aligning treatment intensities with predicted patient outcomes, precision dosing could significantly optimize tDCS therapy and address its inherent variability.

Materials and

Methods: We analyzed a subset of the Augmenting Cognitive Training In Older Adults dataset (NCT02851511) consisting of 333 older adults who underwent tDCS interventions to manage anxiety, using the State-Trait Anxiety Inventory (STAI) State score, dichotomized at 39, to categorize anxiety levels. The dataset primarily consists of electronic health records (EHR), demographic information, and cognitive test scores, offering a detailed overview of patient histories, treatment details, and background characteristics. To predict improvements, we employed Lasso regression, a method that reduces the influence of less significant variables to prevent model overfitting. The dataset was divided into 10 random splits, allocating 80% for the training set and 20% for the test set to ensure diverse exposure during model training. Each split underwent 5-fold cross-validation to optimize the Lasso shrinkage parameter (lambda), which balances model complexity and accuracy. This approach not only minimized prediction errors but also enhanced the reliability and generalizability of our findings, suitable for robust neuromodulation therapy assessments. Additionally, by bootstrapping the model fitting process across multiple data splits, we were able to consistently identify which predictors are influential, ensuring our model’s efficacy in real-world applications.

Results, Conclusions, and Discussions: Our logistic regression models, regularized with Lasso and tailored to the tDCS dataset, demonstrated excellent performance metrics in predicting anxiety improvement. The models averaged an accuracy of 90.75% (95% confidence interval (CI): 88.03% to 93.46%). The F1 score was 63.13% (95% CI: 49.79% to 76.47%), balancing precision and recall effectively. Additionally, the models showcased impressive discriminatory capabilities with a ROC AUC of 92.64% (95% CI: 86.91% to 98.37%). The included figure shows mean coefficient values sorted by magnitude with an ‘occurrences’ heatmap ranging from 1-10. STAI State score at baseline is most predictive of improvement in anxiety with both the highest mean coefficient and presence in all 10 models. This suggests that participants with higher baseline STAI State scores are more likely to experience clinically significant improvements following tDCS therapy compared to those with lower baseline score. Generally, tDCS had a positive impact on anxiety improvement while being present in half of the models trained.

This study validates the use of Lasso-regularized logistic regression to assess tDCS's effectiveness in reducing anxiety, as measured by STAI State scores. Our methods, including extensive cross-validation and bootstrapping, demonstrated strong predictive capabilities, distinguishing effectively between different levels of anxiety improvement. Future work will focus on integrating these predictors with existing brain scan-based tDCS models. This endeavor aims to enhance model accuracy and specificity, promising more personalized and effective neuromodulation therapies that optimize treatment outcomes by leveraging both clinical and neuroimaging data.

Acknowledgements (Optional):