Sensitive & Specific: The Year in AI
Volume 11, Issue 11 | March 13, 2025
ALSO IN THIS ISSUE
Bird Flu update
2024 was the year that AI sucked all the oxygen from the room - mostly in a good way. As the innovations kept coming, we did our best to keep up, focusing on diagnostic-related technologies that we deemed likely to make the biggest impact.
We remain convinced that diagnostics will be the first clinical arena in which AI is broadly used. It is already commonplace in radiologic imaging analysis, and its use in digital pathology as part of image processing (or at least the use of machine learning) is finally expanding. We expect that the next step will be analyzing patient-report data - especially for patients with data from multiple types of testing.
The short-term challenge is proving that the small trials and physician-initiated studies which show exciting results can be scaled and repeatable. The medium- and longer-term challenges are all about physician adoption and reliance on the technology. It reminds us of the early days of genomic analysis - it took physicians many years to believe that this “new” technology could give them more information than they would have gotten from traditional tests and their own examination of the patient. This initial reaction is understandable. It’s a typical wait-and-see approach.
The challenge with AI is the balance between “traditional” human analysis and AI-driven analysis. Will doctors be driven entirely out of diagnosis and monitoring? Of course not. But can AI transform early diagnosis and “see” and predict patterns in individual patients and groups of patients that no human can see? Absolutely yes.
Yes, even with physician adoption, lots of questions on liability, responsibility, and reimbursement remain. But this is a genie that will not (and should not) go back in the bottle.
Here are all those articles in one place, in ten therapeutic categories, for your bookmarking pleasure. (We’ve included the articles from 2025 thus far, too.) You’re welcome.
Cancer
An AI- and spectroscopy-based device that evaluates potentially cancerous skin lesions in folks over 40
A system that can distinguish between positron emission tomography scans of patients with and without lymphoma with 87 - 90% accuracy.
A deep-learning-trained algorithm that can reduce the number of false-positive mammogram results without increasing the number of missed cancers
An algorithm that can identify the originating organ for cancers based on metastatic cells
An AI-based model that can diagnose ductal carcinoma in situ based on chromatin-stained tissue samples
Researchers were able to predict with 98.5% specificity whether people had any of 11 different types of cancer, based solely on changes in their Alu elements (part of what used to be called “junk DNA”
A model that screens electronic health records for patterns that predict pancreatic cancer was able to find 3.5x more cases as compared to current screening protocols
Doctors were 40% better at finding polyps on colonoscopy when they were assisted by AI
An AI-enhanced CT technique was able to detect pancreatic tumors a median of 475 days before clinical diagnosis, with 84% positive predictive value and 90% specificity
Adding AI-based analysis of prostate MRIs to the current standard-of-care screening system identified more men who didn’t need biopsies - without increasing the number of missed cancers
An AI model that looks at how patients’ mammograms change over the course of three years is better than the standard questionnaire at identifying folks at high risk of getting cancer in the next five years
When two radiologists plus an AI model reviewed mammograms, the team caught more cancers than two radiologists without AI did - and didn’t lead to more unnecessary follow-up exams
Cardiology
A free, online app that predicts risk of complications or death after balloon angioplasty +/- stenting
A deep learning model able to diagnose obstructive coronary artery disease based on retina photos
A tool that uses infrared thermal (IRT) camera images to diagnose coronary artery disease
Geriatrics and Aging
A model that uses primary-care data from a patient’s electronic health record to predict their risk of hospitalization after a fall over the next 12 months
Neurology and Mental Health
An AI-based program that predicts whether a patient will progress from amnestic mild cognitive impairment to full-blown Alzheimer’s
Models were able to isolate risk factors for Alzheimer’s disease within electronic medical records with 72 - 81% accuracy
A machine-learning-based diagnostic that uses a short recording of a person’s speech as a sample was able to correctly diagnose moderate to severe depression 69% of the time; when it said a person did not have moderate to severe depression, it was right 75% of the time
Obstetrics and Perinatology
Use of an AI-enabled digital stethoscope allowed clinicians to detect peripartum cardiomyopathy at double the rate of standard care
A fully autonomous AI system that can accurately detect infants who had severe cases of retinopathy of prematurity
A model trained with a deep-learning approach was able to diagnose and rate the severity of pulmonary hypertension in newborns
An AI-based method was able to predict gestational diabetes mellitus as early as 12 weeks
Radiology
AI assistance doesn’t always help radiologists make more accurate diagnoses
Many previously unexploited diagnoses can be extracted from a simple chest X-ray
AI guidance allowed non-specialist ultrasonographers to perform lung exams that, 98% of the time, were just as good as those done by lung ultrasound experts
Respiratory Disease
A machine-learning-based tool that uses CT scans to help diagnose idiopathic pulmonary fibrosis
A novel tool that analyzes speech to detect infectious disease, beginning with COVID and tuberculosis
Sepsis
Use of an AI algorithm that flags patients who may be about to develop sepsis was associated with a 17% relative decrease in sepsis mortality
The FDA has authorized the first AI-based software for the diagnosis of sepsis
Trauma
AI-based methods that use invasive arterial blood pressure, photoplethysmography, and electrocardiography - data that’s routinely obtained in head-trauma patients - to estimate intracranial pressure
Policy
HHS’s health technology regulators published final rules requiring “software vendors to disclose how artificial intelligence tools are trained, developed, and tested.”
The WHO released updated guidelines for large multimodal models (LMMs) used for health purposes
AI in General
Three pieces of essential reading on medical AI for Dx folks
How deep disease insight improves AI for pathology
AI may unlock the diagnostic potential of metabolomics
When AI gets it wrong, check the training set
LLMs in the exam room are still a work in progress
To get AI to recognize rare diseases, teach it what’s common
An AI-enabled tool used short videos to assess vital signs, including blood pressure, pulse, respiratory rate, oxygen saturation, and temperature
Most of the nearly 600 AI / machine-learning enhanced radiology programs and devices aren’t covered by any insurance
Bird Flu Update: Pigs can get the cow strain of H5N1 but don’t transmit to other pigs
H5N1 found in OR wastewater weeks before detection in birds
Good news: A preprint of a study done in piglets indicates that while pigs can get infected with the strain of bird flu found in cows, they don’t spread it to other pigs. In the study, a group of piglets had virus placed in their noses, mouths, and tracheas and were then housed with healthy, uninfected piglets for 14 days. The piglets who were dosed with virus got infected and shed virus particles at a low level. But the rest of the piglets stayed healthy. This work is important because pigs can get infected with both bird flu and human flu, which can cause the two types of virus to share genes.
Avian flu was found in Oregon wastewater six weeks before outbreaks in poultry occurred in that state and seven weeks before the virus was detected in wild birds. The areas where the wastewater was contaminated were near places where large numbers of wild birds congregate.