A Stanford team has taught an algorithm to translate Doppler sounds into clinically meaningful ankle-brachial index ranges—potentially bringing vascular assessment closer to the point of care.
What if assessing limb perfusion began not with squeezing an artery, but with listening to it?
Our colleague and stellar Stanford vascular surgeon Oliver Aalami, working with Adrit Rao and colleagues, has published a fascinating proof-of-concept in npj Cardiovascular Health, a Nature Portfolio journal. Their system, called AutoABI, uses deep learning to estimate ankle-brachial index categories directly from continuous-wave Doppler sounds.
Think of it as a kind of Shazam for limb perfusion: the system listens to a few seconds of arterial music, converts it into a spectrogram, and tries to recognize the hemodynamic pattern.
The problem with our familiar ABI
The ankle-brachial index remains one of the most useful tools for detecting peripheral artery disease. But it is frequently underused because it takes time, requires cuffs and Doppler measurements at several sites, and may be difficult to incorporate into a busy clinic.
More importantly, the ABI can fail precisely in some of the patients in whom we need it most.
In people with diabetes, kidney disease, and medial arterial calcification, tibial arteries may be difficult or impossible to compress. The resulting ankle pressures can be falsely elevated—or simply uninterpretable. The artery has effectively become a rigid pipe, and the cuff can no longer tell us what is happening inside it.
AutoABI approaches the problem from another direction. Instead of asking, “At what pressure does flow return?” it asks, “What can the sound of flow tell us?”
How AutoABI works
The Stanford investigators collected 791 four-second Doppler recordings from 198 patients undergoing formal vascular laboratory testing. An iPhone application used the phone’s built-in microphone, positioned beside the Doppler probe, to record signals from the dorsalis pedis and posterior tibial arteries.
The recordings were converted into time-frequency spectrograms—essentially visual fingerprints of the arterial sounds. Deep-learning models then classified each recording into one of four clinically relevant ABI ranges:
- Less than 0.5
- 0.5–0.7
- 0.7–0.9
- Greater than 0.9
The models performed impressively within the study dataset, with areas under the receiver operating characteristic curve ranging from approximately 0.94 to 0.97 across the four categories.
This is an important distinction: AutoABI does not yet produce a precise numerical ABI. It identifies an ABI neighborhood. But at the point of care, knowing whether a limb appears severely ischemic, moderately impaired, borderline, or relatively normal may be enough to trigger the next appropriate step.
The most intriguing signal: noncompressible arteries
The particularly provocative part of the study involved patients with noncompressible tibial vessels.
Because a conventional ABI cannot provide a trustworthy reference value in this setting, the investigators used independently interpreted waveform phasicity as a surrogate for hemodynamic status. In the small subset of 10 noncompressible recordings, the model’s ABI-range predictions were concordant with the categories expected from the waveform assessments.
That is exciting—but it needs a large asterisk.
Ten recordings are a whisper, not a chorus. The result should be viewed as an intriguing signal of feasibility, not definitive validation. We still need substantially larger, diverse, multicenter cohorts, particularly among people with diabetes, chronic kidney disease, tissue loss, infection, and varying degrees of calcification.
We also need comparisons with toe pressures, toe-brachial indices, transcutaneous oxygen measurements, skin perfusion pressure, pedal acceleration time, and clinically meaningful outcomes such as wound healing and limb preservation.
Why this matters
The larger idea here may be more important than any single performance statistic.
The humble handheld Doppler is already nearly ubiquitous. Its audio output contains information that experienced clinicians learn to interpret—triphasic, biphasic, monophasic—but that interpretation is subjective and varies with training and experience.
AutoABI attempts to turn that subjective sound into a reproducible digital biomarker.
If externally validated, such technology could extend vascular assessment into primary care offices, wound clinics, hospital wards, community screening programs, pharmacies, and eventually the home. It could help identify patients who need formal vascular evaluation before compression therapy, supervised exercise, or treatment of a diabetic foot wound.
This would not eliminate the vascular laboratory. It could make the vascular laboratory more reachable.
The smartphone becomes a bridge: not the final diagnostic destination, but an intelligent front door.
From “toe and flow” to “hear the flow”
For those of us working in limb preservation, the appeal is obvious. Delays in recognizing ischemia remain a major contributor to delayed healing, infection, hospitalization, and preventable amputation. A low-cost tool that helps clinicians identify impaired perfusion earlier could move vascular assessment upstream—closer to the patient and closer to the moment when action matters.
The future of vascular diagnostics may not depend entirely on bigger machines. It may also come from teaching the small devices already in our pockets to listen more carefully.
Sometimes the limb is already telling us the answer.
We simply need better ears.
Reference: Rao A, Battenfield K, Fereydooni A, Chaudhari A, Aalami O. Enabling ankle-brachial index prediction from Doppler sounds using deep learning. npj Cardiovascular Health. 2026;3:21. https://doi.org/10.1038/s44325-026-00116-7
Leave a Reply