Interpretable Clinical Decision Support System for Audiology Based on Predicted Common Audiological Functional Parameters (CAFPAs)

Abstract

Common Audiological Functional Parameters (CAFPAs) were previously introduced as abstract, measurement-independent representation of audiological knowledge, and expert-estimated CAFPAs were shown to be applicable as an interpretable intermediate layer in a clinical decision support system (CDSS). Prediction models for CAFPAs were built based on expert knowledge and one audiological database to allow for data-driven estimation of CAFPAs for new, individual patients for whom no expert-estimated CAFPAs are available. Based on the combination of these components, the current study explores the feasibility of constructing a CDSS which is as interpretable as expert knowledge-based classification and as data-driven as machine learning-based classification. To test this hypothesis, the current study investigated the equivalence in performance of predicted CAFPAs compared to expert-estimated CAFPAs in an audiological classification task, analyzed the importance of different CAFPAs for high and comparable performance, and derived explanations for differences in classified categories. Results show that the combination of predicted CAFPAs and statistical classification enables to build an interpretable but data-driven CDSS. The classification provides good accuracy, with most categories being correctly classified, while some confusions can be explained by the properties of the employed database. This could be improved by including additional databases in the CDSS, which is possible within the presented framework.

Expert validation of prediction models for a clinical decision-support system in audiology

“…The present study clearly demonstrated the importance of combining expert knowledge and statistical learning in the design of a CDSS for audiology. The expert validation and corresponding statistical analyses to investigate agreement between model-predicted and expert-derived CAFPAs provided important insights into the current properties and the necessary future improvement of the CDSS proposed by Buhl ( 22 ) and the prediction models of CAFPAs ( 23 ). Furthermore, the collected qualitative data on the experts' decision-making process are highly valuable to complement statistical conclusions.…”

“…Aiming to further ameliorate clinical applicability, Buhl et al ( 19 – 22 ) and Saak et al ( 23 ) rendered a series of development steps toward a CDSS for audiology, which strongly relies on expert knowledge and is targeted toward future interpretability and integration across different data sources. The CDSS should operate on diverse clinical databases, and it aims at covering the complete audiological decision-making process, including the classification of audiological findings for given patients, as well as suggesting appropriate treatment recommendations (summarized as diagnostic cases).…”

“…The CAFPAs are defined on a continuum ranging on the interval [0 1], with 0 representing “normal” and 1 representing “maximally impaired”. Panel (A) of the figure was taken from Buhl ( 22 ). (B) Schematic representation of the clinical decision-support system (CDSS) by Buhl ( 22 ) (left part, based on labeled data) and relationships to the current study (right, based on unlabeled data).…”

“…Panel (A) of the figure was taken from Buhl ( 22 ). (B) Schematic representation of the clinical decision-support system (CDSS) by Buhl ( 22 ) (left part, based on labeled data) and relationships to the current study (right, based on unlabeled data). Labeled and unlabeled measurement data originate from the same database (light brown box).…”

“…Aiming for further validation of statistically derived CAFPA values, to connect all components, and to finally build a CDSS operable for individual patients (based on labeled data), Buhl ( 22 ) applied the classification approach from Buhl et al ( 21 ) to technically evaluate the predictions in the use case of a CDSS ( Figure 1B , lower left part). The classification was performed on expert-estimated CAFPAs and model-predicted CAFPAs.…”