Nov 17, 2023 · 33m · mad
AI Medical Scribe Breakthrough: DeepScribe CEO Akilesh Bapu on Healing Doctor Burnout
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
On the MAD Podcast, host Matt Turck interviews DeepScribe Founder and CEO Akilesh Bapu to discuss how ambient AI medical scribes are solving physician burnout. They explore DeepScribe's transition from human-in-the-loop data curation to full automation, its multi-model AI architecture, enterprise GTM strategies, and the future of healthcare AI.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 16.5% of the talking time here. How this is scored →
Matt as informed peer 4.1 Guest teaching 3.9 Guest disagreement 0.1 Matt pushing back 0.1
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 25:01 Playful correction regarding Vision Pro preview
The interview is uniformly polite and collaborative; the closest moment to pushback or contrarian framing occurs when the guest corrects the host's belief that a Vision Pro integration preview was real, clarifying it was actually an internal joke.
Hardest push from Matt ▶ 14:52 Host presses on real-time decision mechanics of banditsThe host presses the guest on how the multi-armed bandit algorithm makes real-time decisions during conversations rather than relying on offline benchmarking.
Biggest teaching moment ▶ 13:32 Explaining multi-armed bandit routing over custom speech model buildingThe guest explains how his lead speech engineer corrected his initial instinct to build an in-house speech recognition engine, opting instead for a multi-armed bandit system across multiple API providers.
Matt holds his own ▶ 23:30 Host cites evaluator and fixer model architectureThe host shows impressive research into DeepScribe's internal technical setup by specifically naming their evaluator and fixer model architecture when introducing the topic of hallucinations.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Opening Graphic Animation and Event Title Card | 2 | 3 | 0 | 0 | The host opens with a high-level overview of DeepScribe and asks a broad question about Generative AI adoption in healthcare. The guest answers with background on health system sales hurdles and post-GPT-4 adoption rates. | |
| Navigating AI Burnout and Enterprise Adoption Challenges | 2 | 4 | 0 | 0 | The host asks open-ended conversational questions about healthcare opportunities, challenges, and the guest's background. The guest educates the host on clinician documentation burden and broken data pipelines. | |
| Transitioning from Human-in-the-Loop to Full Automation | 4 | 5 | 0 | 1 | The host demonstrates technical understanding of speech-to-text vs summarization pipelines and asks clarifying questions on multi-armed bandit routing. The guest details the technical setup combining classical models, in-house LLMs, and GPT-4. | |
| Machine Learning Paradigms: Supervised vs. Unsupervised Learning | 5 | 4 | 0 | 0 | The host references articles written by the guest regarding supervised versus unsupervised learning and asks the guest to define them for the audience. The guest explains how both paradigms apply to their model training stages. | |
| Customization Studio and Specialty Clinical Blueprints | 6 | 4 | 0 | 0 | The host demonstrates preparation by citing recent product announcements and specific feature names like progressive notes and blueprints. The guest explains how specialty-specific clinical workflows require distinct data inputs. | |
| Mitigating AI Hallucinations via Evaluator and Fixer Models | 6 | 3 | 1 | 0 | The host demonstrates technical depth by naming DeepScribe's specific evaluator and fixer model architecture to address hallucinations. The guest playfully corrects the host regarding a Vision Pro concept being a designer's joke before discussing go-to-market strategy. | |
| Safeguarding Patient Privacy and Securing Informed Consent | 4 | 4 | 0 | 0 | The host brings up major industry concerns around patient data privacy and third-party LLM data flows. The guest explains informed consent workflows and concludes with his long-term vision for medical error reduction. |
Statements from this episode (16)
DeepScribe initially targeted private practices because large health systems ignored AI startups
“At DeepScribe, a lot of our initial journey was actually through private practices. So we sidestepped health systems because the sales process and the go-to-market Just wasn't there with these large, larger organizations because they didn't take us and other h…”
Bapu: 30% of CMIOs are evaluating and purchasing generative AI
“Now there's a stat floating around that 30% of CMIOs, which is for us, our customer are now looking at generative AI solutions and looking to purchase generative AI solutions.”
Bapu: HLTH conference attendance expanded 20x from 2019 to 2023
“Back in 2019 it was around 500 people in terms of attendance. It was at the MGM Convention Center, which is fairly small. This past year, it was 10,000 people in attendance so literally, 20 x'd in size”
Bapu: Healthcare AI risks triggering another wave of clinician AI burnout
“It also runs the risk of that same burnout occurring again.”
Bapu: Healthcare AI lacks proven system-wide adoption despite strong sales
“People are definitely buying but what's not shown yet is whether these solutions can gain true system-wide adoption.”
DeepScribe CEO: AI achieved parity with human scribes, enabling full automation
“A couple months ago for the first time achieved parity with our AI and our expert human scribes, at which point we took them out of the loop. And so now DeepScribe is fully automated”
Removing humans from the loop requires designing products without safety nets
“In order to get the human on loop, we had to be intentional about it. We had to start building around the fact that a human won't be there anymore.”
DeepScribe ended offshore reviewer contracts as clinicians stopped requesting human review
“Clinicians are barely sending the note to the human to review, yet we were keeping that option. We're keeping this like base level contract with our offshore suppliers. So we decided to officially make the move and that's when we started to see complete shift.”
DeepScribe has gathered over 2.5 million labeled medical conversations
“And we have a little over two and a half million conversations right now. They're all labeled.”
DeepScribe sees more impact fine-tuning in-house LLMs than GPT-4
“So with GPT-IV, to be honest, we haven't gotten the impact we'd like in terms of fine-tuning. Where we've seen the most impact is with our own in-house LLM.”
Bapu: Only 50% of oncologist documentation comes from direct conversations
“But if you look at oncologist, about 50% of their information comes from the conversation. About one fourth comes from a summary that they would have done, and then another fourth may come from previous visits.”
DeepScribe discards LLM output if it lacks classical model validation
“So for a given task, we'll have a classical model typically produce that that same output. And then if the LLM puts in something that the classical model didn't have in its output we'll go ahead and only go with the classical model's output and disregard the L…”
Bapu: Apple Vision Pro adoption in health systems will probably never happen
“It depends on Apple's Vision Pro adoption across its customer base of health systems which probably will never happen, sadly.”
DeepScribe grew from zero to $6M ARR purely through inbound marketing
“We bought AdWords on Facebook and Google and all of our traffic and all of our initial sales from I think zero to six million in ARR came from inbound.”
DeepScribe drafts off Nuance's enterprise market education to win health clients
“We really ride off the backs of Nuance, which is the 800 pound gorilla in our space that is trying to do something similar to what we're doing, but they have the distribution advantage. So they've gone and educated all the customers, and we follow on and basic…”