Artificial Intelligence in Medicine and Public Health: Prospects and Challenges Beyond the Pandemic

Though still in its infancy as a field, artificial intelligence (AI) is poised to transform the practice of medicine and the delivery of healthcare. Powered by breakthroughs in machine learning (ML) algorithms, enhanced computing power, and increasing data volume and storage capacity, AI has made noteworthy advances over the past decade across many medical subspecialties. Experts predict AI-based medical devices and algorithms will play a major role in the delivery of preventive, diagnostic, and therapeutic interventions. In addition to our occasional blog on the topic (see recent example), our Public Health Genomics and Precision Health Knowledge Base (PHGKB) and our weekly update display the latest scientific literature, evidence synthesis, guidelines, evaluation, and implementation studies for the applications of AI in a wide variety of diseases across the lifespan.

A recent Nature Medicine article discusses promising uses of artificial intelligence in medicine, particularly in medical imaging and big data integration, and considers technical and ethical challenges for their applications in improving human health. Here is a quick summary of the review and the implications for population health.

Imaging at the Forefront of AI in Medicine

In the interpretation of medical images — a niche where AI models have made great strides — the AI workflow starts with images that have been read and annotated by human experts. The AI model can analyze and interpret images and compare its interpretation to that of human experts. AI can then learn and refine its interpretation models over time and after analyzing numerous images. AI tools have shown that they can meet, or even exceed, experts’ performance across medical specialties that rely on human interpretation — namely, radiology, pathology, dermatology, gastroenterology, and ophthalmology. For instance, one study used AI methods to analyze whole-slide images and demonstrated that their model was more accurate in predicting patient survival from malignant mesothelioma, compared to current pathology practices. Another study demonstrated that an AI model for the optical diagnosis of colorectal cancer can achieve precision comparable to that of skilled endoscopists. Such advances have demonstrated how AI can refine diagnostic accuracy and improve patient outcome predictions, while enabling a faster clinical workflow and more efficient use of healthcare resources.

Many Promising Avenues for AI in Medicine

In addition to AI for medical image analysis, the review paper presents three promising avenues for AI. First, AI can learn from non-image data sources, such as text and genome sequences, and can broaden the array of possible datasets that can be used for medical insights and drug discovery. AI methods have been used to predict outcomes from medical signal data, such as electroencephalograms (EEG), electrocardiograms, and audio data. For example, AI was recently applied to EEG signals from clinically unresponsive patients with brain injuries to detect brain activity, an important predictor of long-term recovery.  AI can also integrate multiple sources of medical data for improved medical diagnosis. For example, one study for diagnosing respiratory disorders took audio recordings of patients’ coughs as well as reports of their symptoms as input to enhance the diagnosis of respiratory disorders. AI models have also been applied to more complex inputs, such as electronic health records, with various data such as vital signs, prescriptions, and laboratory results.

Second, AI can learn from data without any labels or annotations (e.g. image labelled with a known medical diagnosis), a process known as unsupervised learning. Labeled data can often be costly and time-consuming to obtain. AI advancements that can use poorly labeled data can widen the landscape of applications in medicine. For example, clustering algorithms can organize unlabeled data points by grouping similar data points together have been applied to several conditions such as sepsis, breast cancer, and endometriosis to identify clinically meaningful patient subgroups.

Finally, AI systems that collaborate with human experts can enable a symbiosis between AI and humans to harness the advantages of both and achieve performance that surpasses that of AI or human experts alone. For example, a recent study found that AI-assisted experts surpassed both humans and AI alone when detecting malignancies on chest radiographs. The usefulness of human-AI collaboration will likely depend on specific tasks and clinical scenarios.

AI in Medicine: Technical and Ethical Issues

As discussed in the paper, enthusiasm for the potential of AI technology is accompanied by concerns of data quality, quantity and transparency of AI models, evidence of clinical utility, regulatory challenges, as well as ethical data use and the impact of equity and bias on outputs from AI models. Limited availability and interpretability of large amounts of data, which are key for training AI models, present a pressing practical challenge for implementation. The devices and equipment needed to gather large data are costly and may be unavailable in many health systems, and additionally, hand-assigned annotated datasets by experts can be difficult to attain given time constraints. Another major concern is that for AI models to enter clinical practice, the system must garner user trust. Transparency is key as users must be able to understand the reasoning behind AI prediction models and to recognize potentially incorrect predictions. Another major concern is the ability to account for and remove biases from AI algorithms that may otherwise compromise findings and exacerbate health disparities.

AI and Health: From Medicine to Public Health in the Pandemic Era

The ability of AI models to analyze and interpret large health datasets at scale can also be transformative for public health and epidemiology and lays the foundation for precision public health. The analysis of large health data from many sources related to people, places, and time may provide more and better insights on the determinants of disease on both the personalized and population levels and can help accelerate public health surveillance and shape public health policies and implementation activities.

AI has emerged as a research and public health tool in the response to the COVID-19 pandemic. Table 1 below shows selected recent examples of AI studies in COVID-19, ascertained by our COVID-19 Portal on Genomics and Precision Health (COVID-19 GPH), an integral component of PHGKB. Table 2 shows the numbers of AI/ML studies captured in COVID-19 GPH by category of investigation. While it is too early to assess the impact of AI on the response to the pandemic, these data provide a vivid illustration of the emerging applications of AI in public health.

In conclusion, as discussed in the paper, applications of AI in medicine and public health are still in infancy. While there are no shortcuts on the long road ahead for investigating AI benefits and addressing implementation, ethical and regulatory issues, we are optimistic about the potential for AI to improve the health of individuals and populations.

Table 1

Table 2: Number of Scientific Papers that Use AI/ML in the Response to the COVID-19 Pandemic, by Category of Study

Source: COVID-19 GPH, February 26, 2022
The database was searched using the terms, artificial intelligence, machine learning, and deep learning.
The categories are not mutually exclusive.