Recognition of diseases, in their clinical (diagnosis) or pre-clinical (screening) stage, is an essential process for therapeutic and preventive measures taken to avoid complications or progress to death, as well as to reduce symptoms and ensure the quality of life of patients.72 Both in screening and diagnosis, tests, exams, or other procedures are performed and the results are valued or not, considering the degree of clinical suspicion (pre-test probability), the accuracy of the test, and the diagnostic criteria of each condition clinic. The methods used for the diagnosis and staging of diseases are often important for defining the prognosis of each patient, that is, forecasting the clinical evolution, duration and probable outcome of an individual’s current medical condition, essential for doctors and patients choose the best treatment to be performed, weighing the costs and benefits of the different therapeutic modalities.73
In clinical practice, the diagnostic process involves the complex integration of information obtained from the patient’s history and physical examination, with complementary, imaging and laboratory tests, with the body of information from the medical literature (there are more than 10,000 diseases cataloged) and communication with the patient and his family. Often, such information is integrated into diagnostic or risk scores, used to guide the conduct and the therapeutic plan.13 Part of the information used is obtained by systematic analysis of complementary exams, such as radiographs and cardiological exams, or as a whole, in search of patterns that are typical of a specific condition, or suggestive of an unfavorable course or response to a particular therapeutic measure. The analysis of these exams is usually done by experienced and highly trained specialists, who acquire the ability to recognize such patterns, which are often unrecognizable to other doctors. However, such specialists are not available in all locations and have variable and poorly reproducible accuracy, aspects that are limitations for everyone to have access to diagnosis, especially patients who live in more remote locations. The amount of data that needs to be integrated and analyzed, whether from traditional scans as presented above, or generated by new devices, such as biosensors, wearable devices, and smartphones, typically represents much more than any human specialist is capable of analyzing and processing.
The diagnosis of diseases has been the focus of AI since the mid-twentieth century, when automated diagnostic systems, based on decision rules, were created for different applications, such as the diagnosis of blood-borne infections.14 Such systems have shown promise to accurately diagnose and treat diseases, but, because they are not more accurate and reliable than physicians, they have only been partially adopted in clinical practice. Recent developments in the field of machine learning have enabled algorithms to outperform human performance in tasks such as image classification 74 and speech recognition.75 These advances have inspired great expectations regarding the potential of this health technology76 and the AI-guided diagnosis has been an area of active research with promising results. The use of deep neural networks has also shown promising results in screening and prognosis. Attia et al.77 used convolutional neural networks with 12-lead conventional ECG to track, with high precision (AUC 0.93), the presence of left ventricular systolic dysfunction, a marker of cardiac involvement that can benefit from early pharmacological treatment. The performance of these systems is often limited by the unavailability of data in the required quantity and quality. Brazil, however, has a public health system with universal access and nationwide coverage, with large databases that can be integrated and used in several AI health applications. For example, the Minas Gerais Telehealth Network has a base of ECGs with almost 5 million exams, which has been used by researchers at this Center.9
Artificial intelligence can certainly help to improve the effectiveness of diagnosis, prognosis, and screening, with greater precision, lower cost, and greater reach, expanding access and reducing inequalities in health care. The first challenge, in this case, is related to the volume, diversity, and heterogeneity of data collected over relatively long periods, which are the entry point for models that need to have good accuracy and, at the same time, generate interpretable results for health professionals. Second, the construction and use of these models must respect guidelines of ethics, transparency, privacy, responsibility, explainability, security, reliability, and usability, among others that are relevant to medical practice. Thirdly, it is essential that the validation methodology not only checks the accuracy of the techniques, but captures the understanding on the part of its users. Finally, it is necessary to consider issues of scale and cost of executing these applications, as well as the availability and effectiveness of specific non-invasive devices to collect the necessary data.
For applications that answer all of these technical questions and bring real improvements in medical practice and the results of their activity, their users, for example, doctors, must be able to use them effectively, not just as an auxiliary tool, but as a tool for self-improvement, while contributing to the improvement of models and their applications. Health professional can become increasingly skilled in the task of diagnosis as he acquires experience combined with the knowledge, establishing more quickly and with greater confidence specific disease patterns and relationships. The process of integrating clinical history, symptoms, tests for the recognition of diseases can be greatly favored by AI systems, overcoming limitations and challenges existing in clinical practice. However, he must not pretend to replace human beings in the process, but to integrate and create a virtuous circle between human and machine.
Principal Researchers: Antonio Ribeiro, Marco Romano-Silva, Wagner Meira Jr., Marcos Gonçalves
9. Ribeiro ALP, Paixão GMM, Gomes PR, Ribeiro MH, Ribeiro AH, Canazart JA, et al. Tele-electrocardiography and bigdata: The CODE (Clinical Outcomes in Digital Electrocardiography) study. J Electrocardiol [Internet]. 2019 Sep 7; Available from: http://dx.doi.org/10.1016/j.jelectrocard.2019.09.008
13. Topol E. Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again. Hachette UK; 2019. 400 p.
14. Davenport T, Kalakota R. The potential for artificial intelligence in healthcare. Future Healthc J. 2019 Jun;6(2):94–8.
72. Kosack CS, Page A-L, Klatser PR. A guide to aid the selection of diagnostic tests. Bull World Health Organ. 2017 Sep 1;95(9):639–45.
73. Genetics Home Reference. What is the prognosis of a genetic condition? [Internet]. Genetics Home Reference. [cited 2020 Jun 16]. Available from: https://ghr.nlm.nih.gov/primer/consult/prognosis
74. Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks [Internet]. Vol. 60, Communications of the ACM. 2017. p. 84–90. Available from: http://dx.doi.org/10.1145/3065386
75. Hinton G, Deng L, Yu D, Dahl G, Mohamed A-R, Jaitly N, et al. Deep Neural Networks for Acoustic Modeling in Speech Recognition: The Shared Views of Four Research Groups [Internet]. Vol. 29, IEEE Signal Processing Magazine. 2012. p. 82–97. Available from: http://dx.doi.org/10.1109/msp.2012.2205597
76. Hinton G. Deep Learning-A Technology With the Potential to Transform Health Care. JAMA. 2018 Sep 18;320(11):1101–2.
77. Attia ZI, Noseworthy PA, Lopez-Jimenez F, Asirvatham SJ, Deshmukh AJ, Gersh BJ, 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 Sep 7;394(10201):861–7.