Deep learning models to detect hidden clinical correlates

David Ouyang, James ZouThe Lancet Digital Health 2020 · DOI · OpenAlex

Abstract

It is often said that a picture is worth a thousand words, a statement that is also true in the medical domain. Over the past decade, deep learning algorithms applied to medical diagnostics have displayed the ability to accurately and precisely identify clinical phenotypes and diagnoses beyond their conventional use. From fundoscopic images being used to evaluate blood pressure and cardiovascular risk to echocardiogram videos identifying a patient's age and sex, investigators have shown the ability for deep learning algorithms to capture systemic phenotypes that would be difficult for human experts to evaluate.1Poplin R Varadarajan AV Blumer K et al.Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning.Nat Biomed Eng. 2018; 2: 158-164Crossref PubMed Scopus (759) Google Scholar, 2Gorbhani A Ouyang D Abid A et al.Deep learning interpretation of echocardiograms.Digit Med. 2020; 3: 10Crossref Scopus (148) Google Scholar In The Lancet Digital Health, Joon-myoung Kwon and colleagues applied a deep learning algorithm to electrocardiograms (ECGs) obtained in the emergency department, out-patient department, and from admitted patients, and show good performance in identifying patients with anaemia.3Kwon JM Cho YH Cho SH et al.A deep learning algorithm to detect anaemia with ECGs: a retrospective, multicentre study.Lancet Digital Health. 2020; 2: e358-e367Summary Full Text Full Text PDF Scopus (30) Google Scholar With use of a convolutional neural network, the authors developed a deep learning algorithm with input from ECGs and demographic information to assess a regression task for haemoglobin concentration as well as a binary threshold for anaemia. The work solidifies the claim that we can obtain additional value from the diagnostic testing and imaging that are already routinely obtained in medical care. Previous studies have shown that deep learning algorithms applied to ECGs can identify cardiovascular phenotypes, such as heart failure, hidden arrhythmias, and hypertrophic cardiomyopathy,4Attia ZI Kapa S Lopez-Jimenez F et al.Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram.Nat Med. 2019; 25: 70-74Crossref PubMed Scopus (394) Google Scholar, 5Ko WY Siontis KC Attia ZI et al.Detection of hypertrophic cardiomyopathy using a convolutional neural network-enabled electrocardiogram.J Am Coll Cardiol. 2020; 75: 722-733Crossref PubMed Scopus (95) Google Scholar, 6Attia ZI Noseworthy PA Lopez-Jimenez F et al.An artificial intelligence-enabled ECG algorithm for the identification of patients with atrial fibrillation during sinus rhythm: a retrospective analysis of outcome prediction.Lancet. 2019; 394: 861-867Summary Full Text Full Text PDF PubMed Scopus (461) Google Scholar and the authors extend this direction to show that deep learning on ECGs can predict systemic phenotypes beyond the heart. An important strength of this Article is the testing of the algorithm on an external dataset from another hospital, helping to avoid potential bias, confounding, or model overfitting, which can happen when an algorithm is only tested on an internal test dataset with the same characteristics as the training dataset. Additionally, the authors enhanced the claim of the particular strength of their deep learning algorithm by comparing their model with benchmark models on patient demographics and non-imaging characteristics. Several challenges remain ahead to further understand and deploy artificial intelligence (AI) systems, such as the one proposed by Kwon and colleagues, into routine clinical practice, as model performance and human interactions can vary in real-world settings. Similar to preclinical studies for pharmaceutical therapies, deep learning models are trained in pristine, controlled environments, which might not reflect real-world clinical settings. Real-world studies of AI systems have shown that model performance can be substantially affected by different hardware, seemingly trivial input settings, patient characteristics, and clinical workflows.7Beede E Baylor E Hersch F et al.A human-centered evaluation of a deep learning system deployed in clinics for the detection of diabetic retinopathy. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems.https://dl.acm.org/doi/fullHtml/10.1145/3313831.3376718Date: 2020Date accessed: May 26, 2020Google Scholar Particularly in resource-limited environments, understanding the routine clinical workflow, such as how patients will obtain treatment when obtaining additional diagnostic testing, is needed to establish the true effect of complicated interventions such as AI systems. We call upon the medical AI research community to release the code of their algorithms to allow for the broad testing of such systems in diverse settings to understand their limitations and biases. Ultimately, as with any important medical intervention, randomised controlled trials are needed to fully evaluate the effect of AI systems. Equally important for widespread adoption of AI systems is increasing understanding of their decisions and the ability to interact, troubleshoot, and evaluate the diagnoses, as highlighted by the authors. Previous work has focused on human-interpretable models when applying AI systems for which the input is directly related to the diagnostic task;8De Fauw J Ledsam JR Romera-Paredes B Clinically applicable deep learning for diagnosis and referral in retinal disease.Nature Med. 2018; 24: 1342-1350Crossref PubMed Scopus (1140) Google Scholar, 9Ouyang D He B Ghorbani A et al.Video-based AI for beat-to-beat assessment of cardiac function.Nature. 2020; 580: 252-256Crossref PubMed Scopus (199) Google Scholar however, this is especially challenging when clinicians are asked to evaluate an AI system doing a task beyond the abilities of human experts. The requirement for understanding and trusting the determination of such an AI system is unavoidably higher. Although this question is still open, important corollaries include whether this diagnosis is biologically plausible and are the input data necessary and sufficient to obtain the diagnosis. Designing efficient human-AI interactions is an important and less explored component of this challenge. Given the early stage in development of these algorithms, AI systems are often described in the setting of screening before confirmatory testing by conventional means. However, even incrementally improving the diagnostic yield of a relatively inexpensive routine medical test can have an impact on triage, costs, and the patient experience. The ability to screen for anaemia from ECGs could minimise the need for painful phlebotomy and allow for earlier recognition of the need for blood transfusions. The ongoing development of AI systems for early detection of disease is a promising avenue for obtaining additional value from medical testing. The pursuit of more precise and earlier diagnoses will allow for earlier detection of disease, more efficient triage of health-care resources, and help clinicians make complex decisions with additional diagnostic support. We declare no competing interests. A deep learning algorithm to detect anaemia with ECGs: a retrospective, multicentre studyIn this study, using raw ECG data, a DLA accurately detected anaemia. The application of artificial intelligence to ECGs could enable screening for anaemia. Full-Text PDF Open Access
BibTeX
@inproceedings{ouyang2020deep,
  author = {David Ouyang and James Zou},
  title = {{Deep learning models to detect hidden clinical correlates}},
  booktitle = {The Lancet Digital Health},
  year = {2020},
  doi = {10.1016/s2589-7500(20)30138-2},
}

Loading related content...

Figure