Late Breaking Abstract - Late Breaking Abstract - Supervised Learning with Human Feedback for Enhanced Clinical Decision-Making in Breath-Based VOC Analysis

<bold>Background:</bold> Breath tests for volatile organic compounds (VOCs) hold promise for diagnostics. Yet, challenges persist due to the complexity of VOC origins, influenced by factors like blood solubility and breathing patterns. Machine Learning (ML) applications often act as 'black boxes', complicating clinicians' ability to control confounders and scrutinise findings. <bold>Aim:</bold> We tested whether Supervised Learning with Human Feedback (SLHF) in breath-based VOC analysis can enhance diagnostic accuracy and decision-making by integrating clinician feedback into AI analyses. <bold>Methods:</bold> Conventional ML needs large datasets, which are scarce in pulmonology. We used SLHF to a clinical study (n=67) focusing on 8 VOCs for lung cancer. Explainable AI and human correction techniques ensure transparency, allowing clinicians to adjust for confounders. We trained our initial ML on lung biopsy and GC-MS data and established an analytical baseline using nitric oxide measurements. <bold>Results:</bold> SLHF managed VOC analysis and maintained physician-ML alignment, differentiating between VOC sources and optimising signal-to-noise ratio. Clinician feedback ensured clinician-AI alignment by enabling absolute 30% higher sensitivity (from 67% to 97%) within only a small dataset. Adding the GC-MS data as a training modality to the electrochemical-sensor based ML led to an absolute 74% rise in accuracy (from 17% to 91%) in the detection of biomarkers among electrochemical data. <bold>Conclusions:</bold> Clinicians key role lies in verifying AI outputs, provide feedback, ensure safety, and use AI for shared decision-making. SLHF demonstrates the potential of AI when closely interacting with the clinician.