Conference Program

Explainable Artificial Intelligence for Assessing Impact of Internet-Based Cognitive Behavioral Therapy for Tinnitus (303579)

Gerhard Andersson, Linköping University 
Eldre Beukes, Anglia Ruskin University 
Vinaya Manchaiah, University of Colorado School of Medicine 
*Hansapani Rodrigo, University of Texas Rio Grande Valley 

Keywords: Explainable Artificial Intelligence, Digitial Health, Internet-based cognitive behavioral therapy, SHAP Analysis

There is huge variability in the way that individuals with tinnitus respond to interventions. These experiential variations, together with a range of associated etiologies, contribute to tinnitus being a highly heterogeneous condition. Despite this heterogeneity, a “one size fits all” approach is taken when making management recommendations. Although there are various management approaches, not all are equally effective.

This study aimed to use exploratory data mining techniques (ie, decision tree models) to identify the variables associated with the treatment success of internet-based cognitive behavioral therapy (ICBT) for tinnitus. Individuals (N=228) who underwent ICBT in 3 separate clinical trials were included in this analysis. The primary outcome variable was a reduction of 13 points in tinnitus severity, which was measured by using the Tinnitus Functional Index following the intervention. Analyses were undertaken by using various machine learning algorithms to identify the most influencing variables. In total, 6 decision tree models were implemented, namely the classification and regression tree (CART), C5.0, GB, XGBoost, AdaBoost algorithm and random forest models. As to facilitate explainable artificial intelligence, The Shapley additive explanations framework was applied to the two optimal decision tree models to determine relative predictor importance.Among the six decision tree models, the CART(accuracy: mean 70.7%, SD 2.4%; sensitivity: mean 74%, SD 5.5%;specificity: mean 64%, SD 3.7%; area under the receiver operating characteristic curve [AUC]: mean 0.69, SD 0.001) and gradient boosting (accuracy: mean 71.8%, SD 1.5%; sensitivity: mean 78.3%, SD 2.8%; specificity: 58.7%, SD 4.2%; AUC: mean 0.68,SD 0.02) models we were found to be the best predictive models. Although the other models had acceptable accuracy (range 56.3%-66.7%) and sensitivity (range 68.6%-77.9%), they all had relatively weak specificity (range 31.1%-50%) and AUCs(range 0.52-0.62).