ABSTRACT
Fine-meshed perioperative measurements are offering enormous potential for automatically investigating clinical complications during general anesthesia. In this study, we employed multiple machine learning methods to model perioperative hypoxia and compare their respective capabilities. After exporting and visualizing 620 series of perioperative vital signs, we had ten anesthesiologists annotate the subjective presence and severity of temporary post-intubation oxygen desaturation. We then applied specific clustering and prediction methods on the acquired annotations, and evaluated their performance in comparison to the inter-rater agreement between experts. When reproducing the expert annotations, the sensitivity and specificity of multi-layer neural networks substantially outperformed clustering and simpler threshold-based methods. The achieved performance of our best automated hypoxia models thereby approximately equaled the observed agreement between different medical experts. Furthermore, we deployed our classification methods for processing unlabeled inputs to estimate the incidence of hypoxic episodes in another sizeable patient cohort, which attests to the feasibility of using the approach on a larger scale. We interpret that our machine learning models could be instrumental for computerized observational studies of the clinical determinants of post-intubation oxygen deficiency. Future research might also investigate potential benefits of more advanced preprocessing approaches such as automated feature learning.
