Box 211-19 describes studies that evaluated the clinical use of AI in specialties other than psychiatry.
Box 2
Beyond psychiatry: Using artificial intelligence in other specialties
Ophthalmology. Multiple studies have evaluated using artificial intelligence (AI) to screen for diabetic retinopathy, which is one of the fastest growing causes of blindness worldwide.11 In a recent study, researchers used a deep learning algorithm to automatically detect diabetic retinopathy and diabetic macular edema by analyzing retinal images. It was trained over a dataset of 128,000 images that were evaluated by 3 to 7 ophthalmologists. The algorithm showed high sensitivity and specificity for detecting referable diabetic retinopathy.11
Cardiology. One study looked at training a deep learning algorithm to predict cardiovascular risk based on analysis of retinal fundus images from 284,335 patients. In this study, the algorithm was able to predict a cardiovascular event in the next 5 years with 70% accuracy.12 The results were based on risk factors not previously thought to be quantifiable in retinal images, such as age, gender, smoking status, systolic blood pressure, and major adverse cardiac events.12 Similarly, researchers in the United Kingdom wanted to assess AI’s ability to predict a first cardiovascular event over 10 years by comparing a machine-learning algorithm to current guidelines from the American College of Cardiology, which include age, smoking history, cholesterol levels, and diabetes history.13 The algorithm was applied to data from approximately 82,000 patients known to have a future cardiac event. It was able to significantly improve the accuracy of cardiovascular risk prediction.13
Radiology. Researchers in the Department of Radiology at Thomas Jefferson University Hospital trained 2 convolutional neural networks (CNNs), AlexNet and GoogleNet, on 150 chest X-ray images to diagnose the presence or absence of tuberculosis (TB).14 They found that the CNNs could accurately classify TB on chest X-ray, with an area under the curve of 0.99.14 The best-performing AI model was a combination of the 2 networks, which had an accuracy of 96%.14
Stroke. The ALADIN trial compared an AI algorithm vs 2 trained neuroradiologists for detecting large artery occlusions on 300 CT scans.15 The algorithm had a sensitivity of 97%, a specificity of 52%, and an accuracy of 78%.15
Surgery. AI in the form of surgical robots has been around for many decades. Probably the best-known surgical robot is the da Vinci Surgical System, which was FDA-approved in 2000 for laparoscopic procedures.16 The da Vinci Surgical System functions as an extension of the human surgeon, who controls the device from a nearby console. Researchers at McGill University developed an anesthesia robot called “McSleepy” that can analyze biological information and recognize malfunctions while constantly adapting its own behavior.17
Dermatology. One study compared the use of deep CNNs vs 21 board-certified dermatologists to identify skin cancer on 2,000 biopsy-proven clinical images.18 The CNNs were capable of classifying skin cancer with a level of competence comparable to that of the dermatologists.18
Pathology. One study compared the efficacy of a CNN to that of human pathologists in detecting breast cancer metastasis to lymph nodes on microscopy images.19 The CNN detected 92.4% of the tumors, whereas the pathologists had a sensitivity of 73.2%.19
How can AI be used in psychiatry?
Artificially intelligent technologies have been used in psychiatry for several decades. One of the earliest examples is ELIZA, a computer program published by Professor Joseph Weizenbaum of the Massachusetts Institute of Technology in 1966.20 ELIZA consisted of a language analyzer and a script or a set of rules to improvise around a certain theme; the script DOCTOR was used to simulate a Rogerian psychotherapist.20
The application of AI in psychiatry has come a long way since the pioneering work of Weizenbaum. A recent study examined AI’s ability to distinguish between an individual who had suicidal ideation vs a control group. Machine-learning algorithms were used to evaluate functional MRI scans of 34 participants (17 who had suicidal ideation and 17 controls) to identify certain neural signatures of concepts related to life and death.21 The machine-learning algorithms were able to distinguish between these 2 groups with 91% accuracy. They also were able to distinguish between individuals who attempted suicide and those who did not with 94% accuracy.21
A study from the University of Cincinnati looked at using machine learning and natural language processing to distinguish genuine suicide notes from “fake” suicide notes that had been written by a healthy control group.22 Sixty-six notes were evaluated and categorized by 11 mental health professionals (psychiatrists, social workers, and emergency medicine physicians) and 31 PGY-3 residents. The accuracy of their results was compared with that of 9 machine-learning algorithms.22 The best machine-learning algorithm accurately classified the notes 78% of the time, compared with 63% of the time for the mental health professionals and 49% of the time for the residents.22
Researchers at Vanderbilt University examined using machine learning to predict suicide risk.23 They developed algorithms to scan electronic health records of 5,167 adults, 3,250 of whom had attempted suicide. In a review of the patients’ data from 1 week to 2 years before the attempt, the algorithms looked for certain predictors of suicide attempts, including recurrent depression, psychotic disorder, and substance use. The algorithm was 80% accurate at predicting whether a patient would attempt suicide within the next 2 years, and 84% accurate at predicting an attempt within the next week.23
Continue to: In a prospective study...