Artificial intelligence in medicine: How it works, how it fails

Rick Rejeleene; Neil B. Mehta

doi:10.3949/ccjm.93a.25089

Article Figures & Data

Figures

Tables

Download figure
Open in new tab
Download powerpoint
Figure 1
Different types of artificial intelligence.
Download figure
Open in new tab
Download powerpoint
Figure 2
How artificial intelligence chooses the next word—and makes a diagnosis.

Tables

Figures

View popup

TABLE 1

Architectures of neural networks in machine learning: What can they do?

Architectures	Uses and limitations in healthcare
Supervised
Feedforward neural networks or multilayer perceptrons Data move in 1 direction through layers of “neurons,” each fully connected to the next; the network learns by adjusting the strength of connections	Can be trained on structured patient data (eg, age, comorbidities, vaccination status) to predict a patient’s individual risk of developing community-acquired pneumonia or experiencing a severe outcome, guiding preventive measures or early intervention While powerful, are less specialized for image or sequence data than other architectures
Convolutional neural networks Use filters that scan across the image to detect features like edges, shapes, and textures, the way your eye recognizes parts of a picture	Can be trained to quickly spot signs of pneumonia on a chest radiograph or computed tomography image, determine the extent of lung involvement, or even help assess the severity of the infection by identifying specific patterns, aiding in faster diagnosis and treatment planning Excellent at finding patterns in data that look like a grid, such as images Training data need to include normal studies and other thoracic conditions that mimic pneumonia so that the system can learn to distinguish pneumonia from look-alike abnormalities
Recurrent neural networks Designed to understand sequences of information where order matters Have a memory that allows them to consider past events when processing current information, making them ideal for data that unfold over time	Can monitor a patient’s vital signs (eg, heart rate, respiratory rate), laboratory values, and symptom progression over hours or days Can learn patterns that predict worsening community-acquired pneumonia, like an increasing risk of sepsis or the need for intensive care, enabling earlier intervention
Transformers Excel at understanding complex relationships within long sequences of data (like text), even when important pieces of information are far apart Can process entire sequences at once, efficiently grasping the context and significance of each part	Can read and understand lengthy, unstructured physician’s notes, nursing observations, and discharge summaries related to community-acquired pneumonia and then summarize key findings, extract all comorbidities, or identify subtle mentions of medication side effects, providing a comprehensive patient overview or flagging relevant clinical details for review
Unsupervised
Autoencoders or variational autoencoders Learn to compress complex data into a simplified, meaningful “code” and then reconstruct similar data from that code Are particularly good at finding and representing underlying patterns in data; are often used for data compression, anomaly detection, and generating new data with learned characteristics	Can analyze vast clinical datasets (symptoms, laboratory results, demographics, treatment responses) for patients with community-acquired pneumonia without prior labels Could uncover hidden patient groups (eg, those with a specific inflammatory response that predicts poorer outcomes) that might benefit from targeted therapies, guiding personalized medicine approaches
Generative adversarial networks Composed of 2 competing artificial intelligence networks, one that creates new, realistic data (the “generator”) and another that tries to tell whether the data are real or fake (the “discriminator”); through this competition, they learn to generate highly convincing, novel data	Can create realistic but entirely synthetic computed tomography images showing various stages of pneumonia, which can be invaluable for training other artificial intelligence models when real patient data are scarce or for creating diverse educational materials without compromising patient privacy Could also generate anonymized patient records for research
Supervised or unsupervised
Graph neural networks Designed to understand and learn from data that are structured as a network or graph, where elements (nodes) are connected by relationships (edges); can analyze how information flows and interacts within these complex connections	Could model patient flow within a hospital, identifying high-risk infection transmission pathways for community-acquired pneumonia More broadly, can analyze complex networks of patient conditions, medications, and clinical events, helping to understand how different comorbidities influence community-acquired pneumonia outcomes or how specific antibiotics interact within a patient’s overall drug regimen