In Reply We thank Kar et al for their thoughtful appraisal of artificial intelligence–based decision-support systems and delays between technological advances and clinical application in reference to our article.1 Similar phenomena have been observed for pharmaceutical innovations and evidence-based guidelines, with lag times commonly exceeding 10 years.2 It seems prudent to address these delays in context, promoting early adoption of low-risk, high-benefit interventions supported by level I evidence while exercising caution for high-risk interventions supported by weak evidence. We agree with Kar et al that evidence supporting artificial intelligence augmentation of clinical decision-making is currently weak and discuss 2 potential clinical hazards in a JAMA Surgery article1: the use of biased or misrepresentative model training data could produce erroneous model outputs, and many patients could be harmed in a short time frame if model outputs are not carefully monitored and interpreted. Kar et al wisely propose that large, robust data sets are needed to allow self-learning, minimize the effect of outliers in small data, and ensure model validity and generalizability. We echo these sentiments and remain hopeful that application of the Fast Healthcare Interoperability Resources framework to large-scale, multi-institutional electronic health record data will achieve these important objectives. This would require cooperation and collaboration among clinicians, data scientists, computer engineers, administrators, implementation scientists, and most importantly patients, who deserve data privacy and security as well as personalized, precise, shared decision-making.
It seems likely that biologic limits on the human capacity to assimilate and interpret large volumes of data under time constraints and uncertainty will persist, and that patients will continue to incur preventable harm from diagnostic and judgment errors.3,4 Although further investigation is needed, available evidence suggests that artificial intelligence–based decision-support systems have the potential to mitigate these errors.5,6 Therefore, we maintain guarded optimism that involved parties will continue to work together toward the safe, effective development and clinical application of artificial intelligence systems to augment clinical decision-making.
Acknowledgements
The authors have no relevant conflicts of interest. AB was supported by R01 GM110240 and P50 GM-111152 from the NIGMS. TJL was supported by a post-graduate training grant (T32 GM-008721) in burns, trauma, and perioperative injury from the NIGMS.
References
- 1.Morris ZS, Wooding S, Grant J. The answer is 17 years, what is the question: understanding time lags in translational research. J R Soc Med. 2011;104(12):510–520. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Grant J, Cottrell R, Cluzeau F, Fawcett G. Evaluating “payback” on biomedical research from papers cited in clinical guidelines: applied bibliometric study. BMJ. 2000;320(7242):1107–1111. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Healey MA, Shackford SR, Osler TM, Rogers FB, Burns E. Complications in surgical patients. Arch Surg. 2002;137(5):611–617; discussion 617–618. [DOI] [PubMed] [Google Scholar]
- 4.Shanafelt TD, Balch CM, Bechamps G, et al. Burnout and medical errors among American surgeons. Ann Surg. 2010;251(6):995–1000. [DOI] [PubMed] [Google Scholar]
- 5.Komorowski M, Celi LA, Badawi O, Gordon AC, Faisal AA. The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nat Med. 2018;24(11):1716–1720. [DOI] [PubMed] [Google Scholar]
- 6.Shimabukuro DW, Barton CW, Feldman MD, Mataraso SJ, Das R. Effect of a machine learning-based severe sepsis prediction algorithm on patient survival and hospital length of stay: a randomised clinical trial. BMJ Open Respir Res. 2017;4(1):e000234. [DOI] [PMC free article] [PubMed] [Google Scholar]