250: Canary Speech's CMO Dr. Kang Hsu: What a Former CMIO Learned from the "Dark Side," What 40 Seconds of a Patient's Voice and AI Can Reveal, and Building the ROI Case for Novel AI CDS

Subscribe on: RSS | SPOTIFY | APPLE PODCAST | GOOGLE | BREAKER | ANCHOR

On this episode of The Digital Patient, Dr. Joshua Liu, Co-founder & CEO of SeamlessMD, and colleague, Alan Sardana, chat with Kang Hsu, MD, Chief Medical Officer at Canary Speech, about "What a Former CMIO Learned from the 'Dark Side,' What 40 Seconds of a Patient's Voice and AI Can Reveal, Building the ROI Case for Novel AI CDS, and more..."

Audio:

Guest(s):

• Kang Hsu, MD, Chief Medical Officer at Canary Speech
• Joshua Liu, MD (@joshuapliu), Co-founder & CEO at SeamlessMD

Episode 250 - Show Notes:

[00:06:40] How Kang was "voluntold" into informatics — despite a lifelong love of technology, Kang didn't volunteer when his CNO asked hospitalists for an EHR physician champion; he was assigned. He expected to simply move paper to the computer, but realized the team was reinventing how medicine would be practiced, from order sets with clinical decision support to governance and documentation. As a hospitalist he saw about 15 patients a day; in informatics he could affect "millions of patients being cared for by thousands of physicians." That path led him to an EHR vendor, a CMIO role at Providence, and now CMIO at Canary Speech.

[00:08:59] Why moving from health system to industry feels so different — Kang says health systems move slowly, with so much governance and so many meetings that something that could happen in a month takes two years, while industry is nimble and he reports directly to his CEO. The trade-off is how former CMIOs are received at conferences: with "10 normal attendees and one of us," vendor-side physicians can feel like "the enemy almost" until people realize they've been on the health system side too.

[00:11:59] Why enthusiastic early adopters are a poor measure of readiness at scale — early adopters are essential, but they're tech-savvy and will tolerate friction, working around a bad interface even if they "have to click 19 times and wait for the little clock to spin." They often try to help a project succeed rather than give real feedback. Kang, a self-described naysayer when he started, says the better test is a skeptic, like a physician who still types notes with two fingers and won't try Dragon. If someone like that likes a tool, colleagues will follow.

[00:14:00] What adoption rate counts as success for ambient AI scribes — in Kang's experience, reaching 35% adoption for ambient scribe technology is "amazing," and 50% would be remarkable. The right target depends on the technology: a new EHR needs 100% adoption, while newer tools take time for clinicians to get used to and want to use, especially physicians close to retirement.

[00:18:19] What Kang didn't appreciate about vendors as a CMIO — with vendors "coming out of the woodwork," some of them pushy, he took a defensive stance and tried to stay away when possible. On the vendor side, he's seen that many companies spend years on research, model validation, interoperability, workflow, training, and customer service before commercializing, with CMOs who push back on selling "a half-baked product."

[00:19:54] What Canary Speech's vocal biomarkers can detect — Canary Speech uses AI-powered voice analysis to provide clinical decision support, not diagnosis, for conditions including depression, anxiety, mild cognitive impairment, Alzheimer's disease, Parkinson's disease, multiple sclerosis, and Huntington's disease. It needs as little as 40 seconds of speech, captured through ambient scribes, telehealth, virtual nursing rooms, or wearables. The models ignore the words and analyze about 2,590 vocal elements every 10 milliseconds, or 10 million or more data points versus 150–200 words, producing objective data that patients can't alter the way they can questionnaire answers. About 17 studies are underway, including ADHD, autism, objective pain, PTSD, suicidal ideation, congestive heart failure, and COPD.

[00:23:11] How vocal biomarkers work like a lab value — AI and machine learning made the approach possible, but the idea came from the co-founders' observation that parents can tell when a child who says "I'm fine" isn't, and that some neurologists can hear early Parkinson's in how a patient speaks. Results show that a patient's vocal biomarkers are positive for a condition, to be combined with symptoms, imaging, or labs, and they can sometimes detect disease earlier than symptoms alone.

[00:25:33] Why novelty is the hardest barrier to clinician buy-in — when Canary Speech presented at Providence Grand Rounds, physicians texted Kang more questions than he'd ever received about a presentation. The challenge is convincing clinicians there's another way to screen, using a vocal biomarker profile much like a lab profile or allergy panel. The strongest proof, he says, is the published research showing the models have been validated in more than one forum.

[00:26:39] How vocal biomarkers make the ROI case — the technology is reimbursable through existing CPT codes, and earlier detection can surface HCCs that drive RAF scores and PMPM payments. Many of these conditions go undiagnosed, and earlier treatment can prevent morbidity and downstream utilization. Kang points to depression, which increases non-adherence to treatment for diabetes, hypertension, and CHF, as a source of soft ROI, while noting that "the CFO's looking at the actual bottom line."

[00:28:34] Why more AI-surfaced findings don't have to overwhelm care teams — Kang compares new insights to a patient who mentions "the worst headache of my life" on the way out the door: you'd want to know. Health systems can build follow-up workflows such as a return or video visit, and screening can help prioritize scarce resources. Using a model still in development as an example, he notes children wait four months to a year for a formal autism evaluation, and vocal biomarkers combined with caregiver and clinician input could help decide who gets seen sooner. He also imagines an AI agent one day administering and scoring the 10–15 minute MoCA cognitive test.

[00:32:43] Which CMIOs are likely to thrive in industry — Kang, who lives in Las Vegas, says the move suits people willing to take risks, eager to do something novel, and able to handle people who don't always want to talk to them. It can be a great late-career venture, while early-career physicians are better served by first getting clinical and informatics experience, since companies want someone who can bring clinical vision and strategy.

[00:34:37] Why spontaneous speech beats a scripted reading — reading aloud is different from spontaneous speech, which the brain produces in a different way, and the vocal biomarkers Canary Speech looks for come out more in spontaneous speech. The company captures speech ambiently during visits or, in research, by having patients describe a photo in an app.

[00:35:42] What evidence CMIOs should demand before trusting an AI screening tool — Kang says AI governance committees should ask how a model was created, who was represented in training, and how it performed across age, sex, gender, geography, language, and accent, along with sensitivity, specificity, and false positive rate. CHAI applied model cards put all of this in one place, "basically like the sticker on the back of a Campbell's Soup can," and Canary Speech has published them for its behavioral and cognitive health models. Its behavioral health models used thousands of voice samples from every U.S. state, plus Canada, Australia, and the UK, but health systems should still test in their own populations.

[00:39:21] How health systems should handle consent for voice analysis — Canary Speech recommends consenting patients specifically for vocal biomarker analysis, beyond ambient scribe consent. Wiretap laws vary by state, but Kang says organizations should follow the strictest standard regardless, because "this is like a trust relationship between a patient and a physician." Patients can decline Canary while still allowing the scribe, or opt into only one model, and results appear about 5–10 seconds after the 40-second sample.

[00:41:31] Where prevention ends and surveillance begins — Kang notes that hospitals already label patients with risk-of-violence flags that are often never removed, much like the flag his dog has carried on its vet chart since biting an assistant as a puppy. Canary's safety model scores the room, listening to up to five voices, rather than labeling an individual, and leaves flagging decisions to hospital policy. The 0–100 score combines anxiety and vocal energy, hospitals can tune the alert threshold to reduce false positives, and daily trends show when a room shifts from green to yellow. The goal is to let caregivers "arrive earlier with empathy and not later with backup."

[00:44:52] What part of the physician's job AI should eliminate — clinicians spend an inordinate amount of time clicking, documenting, reconciling, billing, coding, and searching the chart "just to restate something so that the system can recognize that we did something." Kang hopes AI removes this "administrative translation layer," giving physicians more time face to face to provide empathy, build trust, and lead difficult conversations.

Fast 5 Lightning Round:

1. What is your favorite book, or the book you've gifted the most?

Air Conditioning by Hsuan L. Hsu

2. If you could instantly master any skill, what would it be?

"Fluency in any language, or cooking a Michelin-worthy meal with any ingredients"

3. Would you rather have super strength, super speed, or the ability to read people's minds?

"Definitely reading people's minds"

4. What is something in healthcare you believe that others might find insane?

"Clinicians spend way too much time proving care happened rather than actually giving care"

5. What is the last movie or TV show you saw, and what did you think of it?

"Shrinking on Apple TV. One of the best shows I've seen in a long time."

The Digital Patient has been recognized as Feedspot's #1 Patient Engagement Podcast of 2025. Thank you to our listeners for making this happen!