Can Artificial Intelligence Help Reduce Human Medical Errors? – Economic and Policy Implications of AI, Blog Post #4


The following is a summary of Can Artificial Intelligence Help Reduce Human Medical Errors? Two Examples from ICUs in the U.S. and Peru by Miguel Paredes. This paper was presented at the Technology Policy Institute Conference on The Economics and Policy Implications of Artificial Intelligence, February 22, 2018.


Medical errors are prevalent and costly. A 2005 study found 133.3 errors per 1,000 hospitalizations in the United States, costing society tens of billions of dollars. Artificial intelligence, Paredes posits, could help reduce the number and severity of errors by adding large-scale data analysis to human judgement.


Paredes tests his theory with two experiments. The first analyzes how well artificial intelligence predicts the effect of diuretics on medical outcomes for sepsis patients in the United States. The second evaluates the effectiveness of artificial intelligence in predicting child mortality for use in admitting decisions in intensive care units (ICUs) in Peru. Both situations are complicated by the number of individual factors at play, including preexisting conditions and the diagnosis and procedures leading to either sepsis or ICU admittance.


Artificial intelligence, which can combine data culled from specific and related patient files, can identify patterns which may not be immediately obvious to medical professionals. For example, if a patient recovering from a routine operation develops sepsis, artificial intelligence could combine data about his previous health, the state of his immune system, and the type of operation he underwent with data on the effect of diuretics for previous patients with similar symptoms or who underwent similar operations. In contrast, a medical professional may not have as much information readily available, and, in any event, cannot process as much information as quickly.


Paredes’ experiments returned positive results. Artificial intelligence predicted with 78% accuracy the likelihood that a sepsis patient would die after 30-days of ICU discharge. In addition, artificial intelligence predictions of medical outcomes for a test set of children admitted to the ICU in Peru was closer to the observed outcomes than the existing system, a standard prediction score in the medical field.


These results suggest that artificial intelligence may be useful in two scenarios: when not enough data exists to make an informed decision or when so much data exists that a human cannot process it all in a short enough period of time. Particularly in high-stress situations, artificial intelligence might supplement medical professionals’ instincts and allow them to devote more attention to critical thinking and problem solving. In the case of ICU admittance, the results indicate that artificial intelligence might help identify children most in need of care, thus improving allocation of scarce hospital resources. Paredes is quick to note that artificial intelligence should be a complement to, rather than a substitute for, medical knowledge and human intuition. However, if artificial intelligence can process vast amounts of data quickly, keep (near) perfect records, and are unaffected by stress and fatigue, it can significantly improve patient outcomes.