Jul 28, 2017 · 17m · a16z

Yubin Park

0:00 / 0:00

gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions

At the Andreessen Horowitz Academic Roundtable, Dr. Yubin Park, CTO of Accordion Health, presents how data science, machine learning, and healthcare business dynamics intersect to enable personalized healthcare. He demonstrates Accordion Health's 'Apricot' platform, illustrating how predictive modeling can optimize post-discharge patient care and reduce healthcare waste.

Chapters · click to jump

1 statement ▾

Park: Precision medicine is just the tip of the iceberg open →

1 statement ▾

CMS bundled payments hold hospitals accountable 90 days post-surgery open →

3 statements ▾

Park: Naive ML models fail in healthcare due to patient bias open →

Park: Healthcare professionals deeply mistrust the machine learning community open →

Park: Post-discharge optimization fits poorly into standard ML frameworks open →

1 statement ▾

Park: Healthcare provider and payer incentives are completely undocumented open →

1 statement ▾

Park: Personalized healthcare extends to benefit plans and financial incentives open →

1 statement ▾

Park: Crawling scientific literature can model healthcare counterfactuals open →

3 statements ▾

Park: Traditional ML yields more questions than answers in healthcare open →

Park: Doctors require ML outputs to be explainable and scientifically grounded open →

Park: Healthcare requires designing a new ML paradigm from scratch open →

1 statement ▾

Park: Surgeons are performing knee replacements on much sicker patients open →

How this conversation actually went

Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →

The host as informed peer 0.0 Guest teaching 6.4 Guest disagreement 1.4 The host pushing back 0.0

speaking balance: gold is the host, purple is the guest (3 minute bins)

Sharpest disagreement ▶ 3:35 Dismantling naive machine learning assumptions

Park sharply critiques simplified data science approaches in healthcare, arguing that ignoring systemic patient placement biases produces completely wrong conclusions and breeds industry mistrust.

Hardest push from the host ▶ 1:58 Zero host pushback in monologue format

The host does not interject or challenge any claims during the transcript as the episode consists strictly of a presentation by the guest.

Biggest teaching moment ▶ 3:20 Selection bias in joint replacement post-care

Park educates the audience on how sicker patients systematically go to skilled nursing facilities while healthier patients go home, demonstrating why unadjusted machine learning models fail.

The host holds their own ▶ 1:58 No host participation recorded

The host remains entirely silent throughout the presentation transcript, precluding any demonstration of host expertise or pushback.

the scores for every segment, with the reasoning behind each

Topic	Guest teaching	Guest disagreement	Why
Post-Discharge Care Decisions in Joint Replacement	7	2	Park explains the CMS bundled payment model for joint replacements and details why naive machine learning approaches fail due to selection bias. Because this is a monologue presentation, host-side scores are zero.
The Intersection of Medicine, Machine Learning, and Business	6	1	Park outlines his vision for personalized healthcare, emphasizing that understanding system incentives is more complex than completing a PhD. The host does not speak or engage during this segment.
Leveraging Big Data Sources for Healthcare Insights	6	1	Park surveys healthcare big data sources ranging from EHRs to wearables, pointing out that off-the-shelf algorithms yield misleading answers without domain grounding. The host remains silent.
Accordion Health Mission and Industry Collaboration	7	2	Park presents Accordion Health's origin story and demonstrates a joint replacement outcome prediction tool while acknowledging initial industry skepticism toward tech outsiders. Host activity is zero.
Broader Applications and Presentation Conclusion	6	1	Park concludes the talk by extending personalized data models to chronic medication adherence and provider financial incentives. The host does not interject or offer dialogue.

Statements from this episode (12)

Insight

Park: Precision medicine is just the tip of the iceberg

“Precision medicine is actually a very, ah, tip of iceberg, which is, ah, actually a much larger stream of, ah, revolution in healthcare, ah, which is called a personalized healthcare.”

Yubin Park Jul 28, 2017 ▶ 1:38

Assertion Supported

CMS bundled payments hold hospitals accountable 90 days post-surgery

“So in the early summer of this year, CMS, the largest payer in the nation, proposed a new formal bundled payment for joint replacement. Now hostels are accountable for all the events. All the events during the 90 days after knee and hip replacement surgery.”

Yubin Park Jul 28, 2017 ▶ 2:13

Insight

Park: Naive ML models fail in healthcare due to patient bias

“Some people may think this is a very simple machine learning problem. Like, okay, so let's grab some data, analyze it, and then build some sort of model to predict their cost and outcomes, and let's just apply in the reality. Without knowing the fact that the …”

Yubin Park Jul 28, 2017 ▶ 3:24

Assertion Not checkable as stated

Park: Healthcare professionals deeply mistrust the machine learning community

“If you talk to any, ah, any people in the healthcare industry, ah, they have a lot of big, ah, mistrust to the machine learning communities.”

Yubin Park Jul 28, 2017 ▶ 3:57

Insight

Park: Post-discharge optimization fits poorly into standard ML frameworks

“This problem cannot be easily cast into a classification regression problems. Ah, it's not like a traditional machine learning framework.”

Yubin Park Jul 28, 2017 ▶ 5:55

Assertion Not checkable as stated

Park: Healthcare provider and payer incentives are completely undocumented

“Actually, those, all those informations are not very documented. Actually, not documented at all. How they're incentivized, how they work, why they are doing that, all those things.”

Yubin Park Jul 28, 2017 ▶ 7:08

Insight

Park: Personalized healthcare extends to benefit plans and financial incentives

“So what I'm saying here is, ah, every axis, all the components in the healthcare can be personalized. It's not just about the treatment and prevention. It's, ah, it's about, like, service bundle. It's about the benefit plan. It's about the financial incentives…”

Yubin Park Jul 28, 2017 ▶ 9:06

Insight

Park: Crawling scientific literature can model healthcare counterfactuals

“Actually, you can crawl all the scientific articles about because they already performed the cohort studies, A-B testing, all those things. You can analyze all those Of previous studies. And use those things to make to answer the what if questions to the, for …”

Yubin Park Jul 28, 2017 ▶ 10:15

Insight

Park: Traditional ML yields more questions than answers in healthcare

“I tried a lot of those kind of traditional machine learning algorithms to those data sets to solve the real problem to it. Usually, it gives more questions than answers.”

Yubin Park Jul 28, 2017 ▶ 10:43

Insight

Park: Doctors require ML outputs to be explainable and scientifically grounded

“They always want to understand why machine is producing this kind of output to all those things, and it needs to be scientifically grounded.”

Yubin Park Jul 28, 2017 ▶ 11:15

Insight

Park: Healthcare requires designing a new ML paradigm from scratch

“It's more about designing a new paradigm in machine learning from scratch.”

Yubin Park Jul 28, 2017 ▶ 11:35

Assertion Partly supported

Park: Surgeons are performing knee replacements on much sicker patients

“Ah, traditionally, ah, knee replacement was done for only healthy patients, but now actually, ah, they're, ah, being more aggressive They are performing a knee replacement procedures to a very sick patients, too.”

Yubin Park Jul 28, 2017 ▶ 14:15

Yubin Park

How this conversation actually went

Statements from this episode (12)

Turn any episode into a week of clips.