Week 171
Foundation Model, Echocardiography, Histologic Signature, Decompensation, GPT-4, Superhuman AI
In Week #171 of the Doctor Penguin newsletter, the following papers caught our attention:
1. Foundation Model. Foundation models, such as GPT-3, BERT and CLIP, are large AI models trained on massive, diverse datasets that can be adapted to solve numerous downstream tasks. Can the recent advances in foundation model research disrupt the task-specific approach to model development in medical AI?
Inspired directly by foundation models outside medicine, Moor et al. proposed generalist medical AI (GMAI) as a new paradigm for medical AI. GMAI is a class of advanced medical foundation models with 3 key capabilities: (1) solve previously unseen problems by having new tasks explained to them, (2) accept varying combinations of data modalities, and (3) formally represent medical knowledge for medical reasoning. The authors further describe the concrete strategies for achieving this paradigm shift, the potential high-impact applications for GMAI, and the core challenges to address in this perspective article.
Nature
2. Echocardiography. How well does AI perform vs. sonographers in the assessment of cardiac function?
He et al. conducted the first blinded, randomized clinical trial to evaluate the impact on cardiologists' interpretations of left ventricular ejection fraction (LVEF) in echocardiogram studies when given initial assessments made by AI and by experienced sonographers. They found that when cardiologists were given AI's initial assessments of LVEF, they were less likely to make substantial changes in the final assessment than when they were given initial assessments by the sonographers. The AI assessment also reduced the time required for cardiologists to overread and improved their consistency with previous cardiologist assessment. Overall, the AI could give non-inferior and even superior initial LVEF assessments compared to sonographers.
Nature
3. Histologic Signature. Pancreatic cancer currently lacks actionable biomarkers to guide the selection of one chemotherapy regimen over another.
In this study, Nimgaonkar et al. applied AI to whole-slide images of pancreatic tumor resections to identify morphologic features associated with the outcome of adjuvant gemcitabine treatment for pancreatic cancer. They derived a histology-based morphological signature associated with the outcome of patients who underwent resection of pancreatic ductal adenocarcinoma followed by adjuvant gemcitabine. Further, they validated that this signature is likely specific to the treatment rather than a general prognostic marker of the disease process. This signature may help clinicians identify which patients will benefit from gemcitabine-based therapy after resection. It also suggests the consideration of pursuing modalities other than genomics and transcriptomics as potential predictive biomarkers of treatment response, when existing molecular approaches have not been proven to do so.
Cell Reports Medicine
4. Decompensation. There is no standard framework for predicting decompensation for patients presenting in the emergency department (ED) without initial physiologic abnormalities, though unexpected clinical decompensation can arise.
Sundrani et al. developed a multimodal machine learning framework to identify ED patients who will develop tachycardia, hypotension, or hypoxia in a 90-minute window. The model integrates both standard triage data (vital signs, demographics, chief complaint) with features derived from 15min of continuous ECG/PPG waveform recordings, and it predicted new tachycardia with AUROC of 0.836, hypotension with AUROC 0.802, and hypoxia with AUROC 0.713 in a 90-minute window.
npj Digital Medicine
5. GPT-4. Can large language models automatically convert free-text radiology reports into structured formats?
In this study, Adams et al. evaluated the performance of using GPT-4 to convert free-text radiology reports into structured formats. Two board-certified radiologists generated 170 fictitious CT and MRI reports for the evaluation, which cover various body regions and examinations. GPT-4 was able to select the most appropriate structured report template (i.e., head CT template, chest CT template, shoulder MRI template, ...) to use, and it successfully transformed all 170 free-text reports into valid JSON files without error. Besides, it currently only costs about $0.10 to convert a report, and this price may decrease in the future as the technology matures. This proof-of-concept study suggests a solution to structure the vast amount of medical text into structured reporting with minor effort.
Radiology
Read paper
6. Superhuman AI. On March 15, 2016, the AI algorithm AlphaGo defeated a human world champion in Go. What do we know about the impact of superhuman AI systems such as AlphaGo and ChatGPT on human behavior?
Shin et al. analyzed more than 5.8 million decisions made by professional Go players over the past 71 years and compared how humans played Go differently before and after the advent of superhuman AI. They found that novel decisions (i.e., previously unobserved moves) occurred more frequently and became associated with higher decision quality after the advent of superhuman AI. Such an increase in decision quality and novelty could not be fully accounted for by memorization of AI decisions but also by human internalization of the AI’s decision-making logic. These findings illustrate that superhuman AI can encourage novel decision-making by humans in certain domains and suggest that innovative thinking can spread from machines to humans and among humans themselves.
Proceedings of the National Academy of Sciences
-- Emma Chen, Pranav Rajpurkar & Eric Topol