Explainable machine learning predictions to help anesthesiologists prevent hypoxemia during surgery

Abstract

One Sentence Summary: We present a new machine learning based system called Prescience that provides interpretable real-time predictions to help anesthesiologists prevent hypoxemia during surgery.Abstract: Hypoxemia causes serious patient harm, and while anesthesiologists strive to avoid hypoxemia during surgery, anesthesiologists are not reliably able to predict which patients will have intraoperative hypoxemia. Using minute by minute EMR data from fifty thousand surgeries we developed and tested a machine learning based system called Prescience that predicts real-time hypoxemia risk and presents an explanation of factors contributing to that risk during general anesthesia. Prescience improved anesthesiologists' performance when providing interpretable hypoxemia risks with contributing factors. The results suggest that if anesthesiologists currently anticipate 15% of events, then with Prescience assistance they could anticipate 30% of events or an estimated additional 2.4 million annually in the US, a large portion of which may be preventable because they are attributable to modifiable factors. The prediction explanations are broadly consistent with the literature and anesthesiologists' prior knowledge. Prescience can also improve clinical understanding of hypoxemia risk during anesthesia by providing general insights into the exact changes in risk induced by certain patient or procedure characteristics. Making predictions of complex medical machine learning models (such as Prescience) interpretable has broad applicability to other data-driven prediction tasks in medicine.peer-reviewed)

A qualitative research framework for the design of user-centered displays of explanations for machine learning model predictions in healthcare

“…This suggests that model-agnostic, instance-level explanation approaches based on feature influence methods may be a viable approach to explaining model predictions in a way that is both comprehensible and useful to healthcare providers. Although other studies have utilized these approaches to explain predictive models in healthcare, [4,31,40] to the best of our knowledge this is the first study to verify that these explanations would be positively received by healthcare providers.…”

“… Interactive Risk representation Probability Critical care providers should be comfortable with the risk representation format. Risk information in feature influence explanations has been previously reported in terms of odds and probability, [ 31 , 32 ] but provider preferences on these representations are unknown. Odds Explanation display format Force plot Visual representations of risk information may facilitate comprehension of risk [ 33 ].…”

“… Odds Explanation display format Force plot Visual representations of risk information may facilitate comprehension of risk [ 33 ]. Tornado plots and custom visualizations called force plots have been used for feature influence explanations, [ 31 , 32 ] but the effectiveness of these visualizations has not been validated in user studies. Tornado plot …”

“…Critical care providers should be comfortable with the risk representation format. Risk information in feature influence explanations has been previously reported in terms of odds and probability, [31,32] but provider preferences on these representations are unknown.…”

From local explanations to global understanding with explainable AI for trees

“…Explaining predictions from tree models is particularly important in medical applications, where the patterns a model uncovers can be more important than the model's prediction performance [15,16]. To demonstrate TreeExplainer's value, we use three medical datasets, which represent three types of loss functions: 1) Mortality, a dataset with 14,407 individuals and 79 features based on the NHANES I Epidemiologic Followup Study [17], where we model the risk of death over twenty years of followup.…”