Advancing personalized healthcare: leveraging explainable AI for BPPV risk assessment.

Background: Benign Paroxysmal Positional Vertigo (BPPV) is a common vestibular disorder significantly impacting older adults, characterized by brief episodes of vertigo triggered by head movements. Accurate and timely diagnosis of BPPV can be challenging due to its overlapping symptoms with other conditions. Machine learning (ML) offers a promising approach to enhance diagnostic accuracy and efficiency.

Purpose: The primary purpose of this study was to evaluate the efficacy of various ML models in predicting BPPV. This common vestibular disorder significantly impacts older adults. This research sought to build models that could accurately predict BPPV using readily available clinical data and to assess the application of Explainable Artificial Intelligence (XAI) to enhance transparency and trust in ML-driven diagnoses.

Methods: In this study, we trained and evaluated several ML models on a rich dataset from Froedtert Hospital involving 7,760 patients characterized by a diverse range of demographic and clinical features. This robust dataset enabled a detailed exploration of the factors influencing BPPV. By employing explainable AI techniques, we aimed to enhance the predictive accuracy of our models and provide clinicians with transparent and interpretable insights into diagnostic reasoning, bridging the gap between machine learning efficacy and clinical usability.

Results: Gradient Boosting emerged as the most effective model, exhibiting the highest accuracy (85.422%), F1 (0.851), and AUC (0.911). Statistical analysis revealed significant demographic disparities in BPPV occurrence. Specifically, the odds ratio (OR) for BPPV among "White or Caucasian" individuals was 2.433 (p < 0.001), indicating a higher prevalence compared to other races. Conversely, "Black or African American" individuals had an OR of 0.851 (p < 0.05), and "Asian" individuals had an OR of 0.791 (p = 0.26). The study also found an OR of 4.498 (p < 0.001) for "Not Hispanic or Latino" individuals, suggesting a significantly higher prevalence of BPPV in this group. The application of XAI facilitated a deeper understanding and trust in model decisions, particularly highlighting how model predictions align with clinical indicators.

Conclusion: The study confirms that machine learning, complemented by Explainable AI, can effectively predict BPPV with high accuracy and interpretability. Leveraging XAI enhances the usability and acceptance of ML predictions in clinical settings, enabling healthcare providers to integrate these insights into their diagnostic processes. Future work should focus on further integrating these models into clinical practice to facilitate early and accurate BPPV diagnosis.

Supplementary Information: The online version contains supplementary material available at 10.1007/s13755-024-00317-3.