The Dawn of AI-Powered Clinical Decision Support: A Surge of FDA Approvals Signals a New Era in Medical Diagnostics

2018 is set to be a memorable year for AI and the healthcare industry. In February, the U.S. Food and Drug Administration (FDA) successively approved several AI-based diagnostic decision support products. First, on February 14, Viz.AI’s ContaCT application received FDA approval, marking the first AI-based diagnostic decision support product cleared by the agency for stroke. Just one week later, the FDA granted another landmark approval: Cognoa’s deep learning application became the first AI-based diagnostic decision support system approved for pediatric autism. What does this rapid succession of approvals for AI-based diagnostic decision support tools signify? What AI-based diagnostic decision support systems are currently available, and which diseases do they target? This article attempts to answer these questions.

As the name suggests, a Clinical Decision Support System (CDSS) is designed to assist physicians in making diagnostic decisions. This proactive knowledge-based system analyzes at least two types of patient data to provide diagnostic recommendations. Physicians then integrate these recommendations with their professional judgment, thereby enabling faster and more accurate diagnoses.

Depending on their operational modes, existing diagnostic decision support systems are primarily categorized into two types: knowledge-based diagnostic decision support systems and non-knowledge-based diagnostic decision support systems. A knowledge-based diagnostic decision support system comprises a knowledge base, an inference engine, and a user interface. Such systems typically operate based on specific rules, with “if-then” rules being the most common. For example, if certain symptoms are present, the system recommends the use of Drug A and provides corresponding medication warnings. Due to their reliance on rules and a knowledge base, these systems require a certain degree of manual intervention.

Non-knowledge-based diagnostic decision support systems primarily leverage AI deep learning. By learning from historical case data, AI algorithms automatically provide recommendations by comparing them with data from current patients. Depending on the implementation method, these systems are categorized into three models: machine vectors, artificial neural networks, and genetic algorithms. The accuracy of such diagnostic decision support systems depends on the algorithm and the size of the training dataset; with superior algorithms and sufficient samples, high accuracy can be achieved. However, due to algorithmic limitations, these systems are only effective for specific diseases.

The FDA regulates medical devices in three classes based on their intended use and the risk they pose to patients. Class I includes low-risk devices, such as medical gloves; Class II comprises moderate-risk devices, such as CT scanners; and Class III consists of high-risk devices, such as stents.

There are two types of AI imaging systems: computer-aided detection (CADe) and computer-assisted diagnosis (CADx). CADe is used to detect abnormalities, while CADx assesses the presence of disease, such as its severity, classification, or prognosis.

The FDA has extensive experience in regulating CADe software and provides 510(k) guidance standards on how to conduct clinical performance evaluations. However, the FDA has historically classified CADx systems as Class III devices.

(510(k) submissions are premarket applications filed with the FDA to demonstrate that the device seeking market clearance is as safe and effective as a legally marketed device not subject to Premarket Approval (PMA), i.e., substantially equivalent. Applicants must compare the device seeking clearance to one or more similar devices already on the U.S. market, and provide evidence supporting the conclusion of substantial equivalence.)

However, in September 2017, the FDA just approved a breast CADx product as Class II, requiring only a 510(k) submission. In other words, the approval was obtained by lowering the regulatory threshold.

The FDA stated that it is currently establishing a regulatory framework for similar clinical decision support systems, with the aim of helping system developers determine the most appropriate treatment plans for patients’ disease conditions. The FDA expressed its hope to encourage developers to create, adjust, and expand the functionalities of such software.

Viz.AI’s ContaCT is the first FDA-approved AI diagnostic decision support system for stroke. According to statistics from the U.S. Centers for Disease Control and Prevention, stroke is currently the fifth leading cause of death in the United States, claiming approximately 795,000 lives annually, and it is also one of the primary causes of disability among adults. ContaCT analyzes brain CT images of stroke patients to identify the CT imaging patterns most closely associated with stroke.

Once new brain CT images are found to match the established pattern, indicating a likelihood of large vessel occlusion (LVO), the system automatically sends an alert report to the physician. By integrating with CT scanners, it can extract CT images in real time and transmit them to tablets or smartphones used by physicians, offering greater efficiency compared to the traditional workflow that requires waiting for CT film development. In clinical trials, ContaCT was evaluated on 300 CT images for the detection of cerebral large vessel occlusions and compared against two experienced radiologists. It achieved an area under the curve (AUC) of 0.91, with both sensitivity and specificity for identifying large vessel occlusions reaching 90%.

Meanwhile, it can notify physicians more rapidly, with the time from CT image generation to final physician alert being less than 6 minutes, significantly accelerating diagnostic speed. Depending on the scenario, ContaCT can save approximately 6 to 206 minutes, with an average time saving of 52 minutes.

In July 2017, the FDA approved Cardiolog Technologies’ electrocardiogram (ECG) analysis platform. This technology is a cloud-based cardiac monitoring and analytics web service designed to assist physicians in screening for symptoms of atrial fibrillation and other arrhythmias using long-term ambulatory ECG monitoring records.

Recently, Lepu Medical’s self-developed “AI-ECG Platform Artificial Intelligence Automatic Analysis and Diagnosis System for Electrocardiograms” was also accepted by the FDA.

Arterys’ Cardio DL is the first FDA-approved artificial intelligence diagnostic decision support system based on deep learning, having received FDA clearance as early as January 5, 2017. This platform leverages convolutional neural network algorithms to learn from magnetic resonance imaging (MRI) scans of at least 1,000 cases, identifying 10 million associated patterns to enhance the recognition of cardiac pathologies. Cardio DL requires only 15 seconds to analyze each case and render a diagnosis, significantly reducing the time compared to the 30 minutes to one hour typically needed by specialist physicians.

Meanwhile, advancements in Cardio DL’s security mechanisms have alleviated concerns regarding the risk of patient privacy breaches. When medical images are uploaded to Cardio DL’s cloud platform, they consist solely of imaging data without any user-identifiable information. Only authorized physicians have the permission to synthesize and reconstruct these images with patients’ personal information for output.

Leveraging its experience with Cardio DL in cardiology, Arterys has also launched AI products for pulmonary and hepatic diseases, all of which have received FDA clearance.

However, according to industry insiders, the product certified by Arterys is not an AI system capable of providing diagnostic results. What was certified is a workstation designed to enhance the workflow efficiency of technologists.

On February 21, 2018, the FDA announced the clearance of Cognoa’s eponymous app, marking the first AI-based diagnostic decision support system for autism spectrum disorder in children. As readers may know, a phenotype refers to the observable physiological or biochemical characteristics of an organism, such as humans.

For young children, a behavioral assessment lasting several hours is typically required to determine whether their development is proceeding normally. These assessments must be conducted by qualified, trained physicians who can analyze developmental progress through interviews with parents or the child. Securing an appointment for such evaluations is difficult, and on average, children do not receive this specialist assessment until the age of 4.1 years. Unfortunately, by this time, they have already missed the window of peak brain plasticity, during which interventions are most effective.

According to research on young children with autism conducted by The New England Center for Children, the effectiveness of 20 to 30 hours per week of one-on-one therapy varies significantly across different age groups, with age two serving as a critical threshold. Ninety percent of children aged two and under achieved “significant improvements” in social and communication skills, whereas only 30% of children who began therapy at age 2.5 or older attained such “significant improvements.” Although Cognoa’s application appears simple, it is powered by an artificial intelligence platform with self-learning capabilities.

By studying and analyzing authoritative autism databases, including the National Database for Autism Research funded by the U.S. National Institutes of Health, Cognoa has summarized behavioral patterns associated with autism. The app asks parents 15 questions, representing the minimum viable number of behaviors needed to indicate whether a child is likely to have autism. Currently, the youngest age at which Cognoa has identified autism in children is just 18 months, thereby securing more treatment opportunities for potential pediatric autism patients.

IDx, a US-based company, recently announced that the FDA has expedited the review of its AI-powered diagnostic decision support product, IDx-DR, with certification expected to be granted soon. This AI system is designed to detect diabetic retinopathy, the leading cause of blindness among patients with diabetes. It is intended to operate without the assistance of ophthalmology specialists, which could have a significant impact on patient care.

Currently, patients often need to wait weeks or months to see an ophthalmologist, which may lead to blindness due to delayed diagnosis. If approved by the FDA, IDx-DR is poised to become the first AI diagnostic system deployed at the frontline of healthcare. IDx holds multiple patents related to optical coherence tomography (OCT), including a patent for a method that automatically assesses vision loss from glaucoma using OCT. Glaucoma is the second leading cause of blindness worldwide and affects more than 3 million Americans.

In addition, IDx has also developed algorithms for detecting macular degeneration and assessing the risks of Alzheimer’s disease, cardiovascular disease, and stroke.

The collaboration between Memorial Sloan Kettering Cancer Center (MSKCC), a leader in the field of oncology, and IBM, a pioneer in artificial intelligence, gave rise to Watson for Oncology. This AI system, developed by IBM, was trained over four and a half years by experts at MSKCC. It has assimilated knowledge from 3,469 medical textbooks, 248,000 journal articles, 69 treatment regimens, 61,540 experimental data sets, and 106,000 clinical reports, while also incorporating the clinical practice guidelines issued by the National Comprehensive Cancer Network (NCCN). It provides decision support for major cancers, including gastric cancer, lung cancer, rectal cancer, colon cancer, breast cancer, and cervical cancer. In clinical trials, it took an average of 24 minutes to screen for breast and lung cancer options, which is 78% faster than the average 110 minutes required by physicians.

By extracting key information from patients’ medical records, it can generate multiple treatment plans, select the optimal diagnostic recommendation to present to physicians, and provide detailed rationale and explanations. According to a study conducted in India, the treatment recommendations provided by Watson for Oncology aligned with physicians’ plans in 96% of lung cancer cases; for rectal and colon cancers, the concordance rates with physicians’ recommendations were 93% and 81%, respectively.

Renowned deep learning scholar Andrew Ng and his team at Stanford University proposed a new technology called CheXNet. This 121-layer convolutional neural network algorithm, known as CheXNet, was trained on ChestX-ray14, currently the largest open-access chest X-ray dataset. The ChestX-ray14 dataset comprises 100,000 frontal-view X-ray images covering 14 different diseases. In comparison with four radiologists possessing 4, 7, 25, and 28 years of professional experience respectively, CheXNet surpassed the previous state-of-the-art performance by at least 5% in detecting masses, nodules, pneumonia, pneumothorax, and emphysema.

In June 2017, collaborators from Harvard University, Massachusetts General Hospital, and Huazhong University of Science and Technology designed a program that combined fMRI brain scans with clinical data to predict Alzheimer’s disease. The team developed a deep learning framework aimed at integrating fMRI brain scans with deep learning techniques. As with any fMRI scan, the images revealed the locations of neural activity in the brain and how these regions were interconnected.

The team began with data from patients with mild cognitive impairment (MCI) and 101 healthy controls from the Alzheimer’s Disease Neuroimaging Initiative. Based on time-series data from 130 functional magnetic resonance imaging (fMRI) measurements across 90 brain regions in participants, researchers were able to identify where signals flickered over time. Using deep learning, AI can interpret the intensity of these patterns and combine them with clinical data on age, sex, and genetic risk factors to predict whether an individual will develop Alzheimer’s disease. The research team reported an accuracy rate of up to 90%.

Yasen, in collaboration with Beijing Xuanwu Hospital, Peking University People’s Hospital, and Peking Union Medical College Hospital, has launched a multimodal artificial intelligence product for brain function assessment. By analyzing data from MRI, PET, SPECT, EEG, and other modalities, this system can be applied to the quantitative analysis, diagnosis, and prediction of various brain functional disorders, including Alzheimer’s disease, epilepsy, Parkinson’s disease, and hemophagocytic syndrome. As of October 2017, the system had been deployed in more than 30 large Grade IIIA hospitals across China, completing cumulative case analyses for over 7,000 patients, with an average accuracy rate exceeding 84% across various disease categories. This system is also the first AI platform in China to provide multimodal analysis tailored to specific diseases, marking a significant step forward in assisting clinicians with diagnosis and treatment.

A research team from Sun Yat-sen University and Xidian University collaborated to develop CC-Cruiser, an AI program capable of diagnosing congenital cataracts. Leveraging deep learning algorithms, the system predicts disease severity and provides treatment recommendations. In computer simulations, the AI distinguished between patients and healthy individuals with an accuracy of 98.87%. Furthermore, it achieved accuracies exceeding 93% across three key metrics: lens opacity area, density, and location.

Moreover, the accuracy of the treatment recommendations provided by the system reached 97.56%. In clinical trials, CC-Cruiser utilized 57 pediatric ocular images from three partner hospitals in China. The recognition accuracy reached 98.25%; it exceeded 92% for all three severity factors; and the accuracy of treatment recommendations was 92.86%.

Stanford University developed a deep learning algorithm based on the GoogleNet Inception v3 architecture, which is a convolutional neural network. Stanford researchers then fine-tuned this algorithm by collecting 129,000 images from 2,000 different skin cancer cases, creating the largest dataset used for skin cancer classification. In this study, the algorithm was evaluated against 21 board-certified dermatologists. The physicians examined hundreds of images of skin lesions to determine whether they would recommend further testing or reassure patients that the lesions were benign.

The algorithm reviews the same images and provides diagnostic results. The AI’s performance is consistent with that of experts. For example, the program can distinguish keratinocyte carcinoma—the most common form of human skin cancer—from benign skin growths known as seborrheic keratoses.

It is evident that current AI-enabled diagnostic decision support is predominantly concentrated in the field of medical imaging. This sector has witnessed the most rapid growth, as an increasing number of diagnoses rely on medical imaging to reveal internal pathological changes. Statistics show that the annual growth rate of medical imaging data in the United States is 63%, while the number of radiologists is growing at only 2% per year. In China, the annual growth rates for medical imaging data and radiologists are 30% and 4.1%, respectively.

On one hand, radiologists hold a relatively low status within hospitals. On the other hand, the mismatch between imaging demand and the number of physicians has led to excessive workloads for doctors, thereby compromising diagnostic accuracy. Leveraging AI for image interpretation and assisted diagnosis can enhance efficiency and effectively bridge this gap. Meanwhile, deep learning-based computer vision is inherently well-suited for addressing imaging challenges, and recent significant advancements in both hardware and algorithms make it unsurprising that most AI-powered clinical decision support systems are concentrated in this domain.

However, AI is currently better suited for solving tasks with well-defined parameters. Consequently, at present, AI-based diagnostic decision support systems are far more mature in medical domains characterized by “hard” data (such as pathological images) than in those relying on “soft” data (such as general diagnoses derived from electronic medical records).

Since the U.S. Food and Drug Administration (FDA) established its Artificial Intelligence and Digital Health Review Division in 2017, AI-based medical products have received substantial support from the agency. In the first three months of 2018 alone, multiple AI-powered diagnostic decision support systems were successively approved—an unprecedented development in the FDA’s history—clearly signaling the strong trend toward the integration of artificial intelligence into healthcare. This trend encompasses not only the currently prominent field of image recognition but also approaches such as radiomics, which integrate imaging data with clinical information for comprehensive computational analysis.

In addition to the United States, Chinese regulatory authorities are also actively researching approval mechanisms for medical artificial intelligence products. On September 4, 2017, the China Food and Drug Administration (CFDA) issued a new version of the "Medical Device Classification Catalog," adding categories corresponding to AI-assisted diagnosis.

According to the latest classification regulations, if diagnostic software provides diagnostic recommendations through algorithms and serves only as an auxiliary diagnostic tool without directly issuing diagnostic conclusions, it shall be registered as a Class II medical device. However, if it automatically identifies lesions and provides explicit diagnostic prompts, it shall be regulated as a Class III medical device.

In addition, the National Institutes for Food and Drug Control (NIFDC) is also formulating detailed approval guidelines. An expert team for fundus image calibration has already been established, and approval standards for pulmonary nodules are currently under development.

AI will bring profound changes to future medical technologies and serve as a powerful driver for innovation and reform in medicine.