🎙️ #18 Andy Beck — CEO and Co-Founder of PathAI

Welcome back to The Good AI Pod! We had a chance to sit down with Andy Beck, to talk through his journey starting and the potential for doing good in healthtech and biotech companies! We’re excited to share our first episode with video 🎥 !

Of course if you’re interested in the usual audio podcast you can subscribe and find it on Spotify, Apple Podcast, or your favorite podcast platform! Check out the quick summary below if you want the TLDR;

Background

[0:30] Andy trained as a physician and pathologist (diagnosing diseases via tissue samples) and got his PhD in ML at the same time. He did research before starting PathAI in 2016 to improve patient outcomes with AI-powered pathology.

[1:50] PathAI, like any startup, faced many ups and downs:

• Building technology that’s better than the rest. PathAI rode the Deep Learning wave in 2016 to surpass existing diagnosis solutions using SOTA ResNet architectures and won the CAMELYON 2016 challenge, but still had to work to platformize the technology.
• Selling and delivering value for biopharma customers in their translational research and clinical trials required good relationships and trust with biopharma teams — Andy had developed many of the relationships doing research.

[4:19] PathAI overcame the initial cold start problem of data in their first project by adhering to the lean startup method and partnering with a large biopharma to build and learn iteratively:

• This biopharma provided a small set of gigapixel-sized whole slide images after PathAI gave them a discount and promised to deliver AI biomarkers.
• Annotations were the real bottleneck on these 100,000x100,000px images. In the early days Andy annotated whole slide images himself!!!
• Today PathAI can get thousands of annotations in hours with their network of thousands of annotators and collect data from multiple third-party and customer sources.

PathAI Business Model

[8:31] Sell accurate and reproducible analysis of gigapixel images to enable high quality research for improved clinical outcomes and higher quality clinical care.

• Initial investment in building the foundational ML models might be high but it is amortized over time with the right unit economics.
• Over time the models improve iteratively, improving the customer experience
• This is sustainable because this is a critical need for research and clinical practice and the accuracy, reproducibility and predictive power of the models are tied to the financial success of the business.

[11:27] What is the common thread between research and clinical practice?

• Similar to foundational models like GPT where predicting the next word is the core, at PathAI, the core unit of analysis is a gigapixel whole slide image of tissue, segmentation of those tissues and classification of cells — this basic functionality underlies all of their products.
• Downstream human interpretable features can be calculated downstream for research purposes but those same features and embeddings can be used for clinical diagnosis as well.
• Regulatory oversight may be different between and within clinical trials and clinical practice. PathAI focuses on finding common regulatory requirements and ensuring high adherence to those across both types of deployments.

Balancing Shareholder Value and Mission

[13:49] PathAI is in the business of accurate, reproducible and predictive diagnoses which naturally are tied to improving patient outcomes.

• Improved diagnostics result in better therapeutics which increase survival rates of patients and are tied to financial success of the business.
• PathAI has no incentive to falsify or fake results adversarially since that would reduce the likelihood of the customer being able to find an effective drug (and thus the likelihood of them continuing to pay) and so the customer, shareholders and PathAI team are all aligned on patient outcomes.

[15:16] PathAI focuses on clinical care where the incentive alignment is clearer than administrative tasks in biopharma like advertising or sales. The stakes are high because the patient is at the center of every interaction and pharma could be risking lives and $100Ms if data didn’t meet quality standards.

PathAI Solves the Dearth of Expert Diagnosticians

[18:05] PathAI enables expert diagnosis at scale through its human-in-the-loop AI making existing pathologists more effective.

• Currently PathAI is available in the US but hopes to expand to places where this shortage is more acute.
• Imagine a pre-screen of every patient to enable pathologists to make decisions that surpass even a consensus of pathologists. This future would bring the expertise of multiple pathologists to every patient!

Why For-Profit vs. Non-Profit?

[21:02] Andy worked for many years in non-profit research organizations such as Harvard Medical School and Beth Israel and found that bringing new ideas to market was a slow burn.

• For-profit organizations can raise money from investors looking for a return to quickly iterate, attract talent and deliver results to customers with all stakeholders aligned on the goal.
• Inspired by Apple, Google, Microsoft, Amazon which were all relatively recent tech startups which were privately funded. The model is tried and true.

[25:15] The vision for companies is to become self-sustaining so you can reinvest in the company and not rely solely on fundraising as the path to sustainability.

We switched gears around halfway to talk more about the general healthcare and biotech space, the opportunities and the challenges. We referenced Bessemer Venture Partners 2023 Healthcare and Life Sciences Predictions for some of these topics (highly suggest reading through it if you’re interested in building in the space :)

Healthcare and Biotech AI a Net Good?

[26:40] Companies in the healthcare and biotech spaces fit multiple molds — if they do something bad, the consequences can be much worse given the direct impact to patients:

• Good: diagnostics or therapeutics companies focused on curing diseases (especially terminal) for patients are nearly always doing something positive.
• Bad: companies providing administrative tasks that are played to further bloat the system vs. reduce costs or improve outcomes (e.g. billing bloat, targeting ads randomly without outcomes in mind, etc)
• Bad: companies who had noble missions but founders or operators did something fraudulent or illegal demonstrating a lack of integrity (e.g. money laundering)

[29:30] Outside of patient outcomes, there are many tasks in between that could also be good (e.g. identifying the right tests, increasing efficacy of drugs, etc.) Keeping focused on how the in-between tasks ultimately help patients is helpful to ensure companies don’t stray from or turn a blind eye to a potential negative effect of their products.

Ballooning Healthcare Costs

[32:01] Biopharma spends billions on research & development which have paid off for many diseases such as cancer, obesity, cardiovascular disease, etc. However,

• our current systems incentivizes drug development at the expense of other types of coaching therapy,
• more research should be done on changing behaviors since that can also be a cheaper way to achieve similar outcomes and
• payers and regulators also have a disproportionate say in the pricing of drugs which often drive up the cost.

[36:24] How do we get out of this loop?

• AI tools can help with efficiency and reduce administrative spend.
• Focusing on critical unmet needs for patients whe developing drugs.
• Less admin spend can increase the % of dollars spent on understanding disease biology, development of therapies and getting effective therapies approved.

Generative AI Opportunities in Healthcare

[39:06] PathAI can leverage generative approaches in multiple ways

• Synthetic data generation of WSI patches with different technical artifacts, image, tissue and stain quality.
• Communication of image data, the literature and key results to the patient, treating oncologists, surgeons and folks of all educational and socioeconomic backgrounds. Synthesis, retrieval and summarization of literature, patient history, image information can help make the results of these tests actionable for clinicians and patients.

[41:06] Can we use language models to help craft doctor’s notes or referral notes to specialists?

• Definitely, companies like Open Evidence are crawling the biomedical literature and summarizing results with citations.
• Bard already leverages LLMs on top of a search graph of websites and similar approaches can be used for the more narrow medical domain.
• Leveraging LLMs coupled with traditional retrieval to do deep literature search and mapping to a patient profile with residents and physicians as a last mile to interpret and make decisions (e.g. Retrieval Augmented Generation).

[43:23] Financial incentives are important to consider to help adopt these generative language models. In healthcare, a few of these include:

• Notes are used to determine billing codes, so the more accurate they are the better for hospital systems
• Physician efficiency can help reduce physician burnout and ideally improve patient care, but incentives are always difficult to align.

[45:55] Bot Wars!!! Payers and Providers may both be armed with LLMs to help either bill more or fend against incorrect billings so it’s unclear how it may shake out.

• Andy sees providers thinking about this more perhaps because of thin margins, while payers may start to look into this only once they feel they are paying too much.

Inflation Reduction Act Impact on Biopharma

[48:05] Pharma had to adjust their portfolios to deal with the reduction in profits from the reduced patent shelf life of small molecule drugs. It’s not clear if the drugs that had to be shelved due to the regulation will have an adverse impact.

Biosecurity Risks Due to AI Drug Discovery?

Much like other AI technologies, is it possible for AI-based drug discovery to enable bad actors to develop biological weapons? And how do we prevent it?

[49:49] Good area for government or an international consortium to put forth standards for risk mitigation to prevent biosafety issues.

• Startups could start here much like cybersecurity, but unclear how that might play out in the long term given biosecurity isn’t an existential threat for every company (but could be for governments)

Advice for Future Founders

[51:44] Build Startups!!! And focus on a few key areas

• Start with the question: “What are the biggest challenges? What are the pain points and where are the dollars being allocated?” NOT with the technology.
• Invest in understanding the customers deeply, their sales cycles, obstacles to adoption and financial ROI. Don’t start with a deep technology and hope you find a good use case…it’s an uphill battle.

Hope you enjoyed this episode!!! Here are the video and podcast links again :)

If you liked this and want to hear more from AI leaders and company builders focused on a positive mission, please subscribe or reach out if there’s someone you’d like to hear from!!!

Thanks and catch you on the next one :)

✌🏾Anand

Originally published at https://thegoodaipod.substack.com.