AI/ML Advice for HealthTech and MedTech Start-ups
December 2023 + January 2024
This article has been adapted from advice I shared on the “Quroba Conversations Artificial Intelligence and Machine Learning in MedTech” panel, with Brian Dolan, Gary Urban, and Kwame Ulmer, moderated by Evan Tsang, in November 2023. Here is the:
Full recording of the session;
30-second snippet where I share the #1 mistake I see HealthTech start-ups make when developing ML/AI models for diagnosis.
Common Mistakes see HealthTech and MedTech Starts-up Make
Trying to disrupt an industry you don’t have expertise in, whose nuances you don’t understand. Potential side effects of this will include a lack of adaption (potentially because the solution doesn't tie into existing workflows) and data leakage.
For example, Epic's AI algorithm to predict sepsis (~2017) suffered from data leakage (more on this below), because it used a physician's order for antibiotics, which were prescribed by the physician to try to treat sepsis, as one of the indicators of sepsis.
Another example of data leakage is an ML model predicting the existence of a condition from MRI images; this model will also seem to do well because to get MRI images, a patient needs a referral to a specialist and then be seen by a radiologist who will capture any “weird-looking” images. This way, a naive model that predicts everyone has the specific condition – will do very well (!).
TIP: As Karen Fischbach of Point32Health advised start-ups at Techstars Healthcare Demo Day in June 2023, try to understand the relationship between the 3Ps: provider, patient, and payer.
Trying to disrupt workflows and status quo you don’t understand.
You don’t want to be a CEO on “Undercover Boss”, suggesting improving KPIs you don’t understand (!)
TIP: I highly recommend having clinical and technical expertise in HealthTech/MedTech to walk you through the nuances of the US Healthcare System from the clinical and infrastructure side, to help you understand, the following, and help formulate and answer the business question:
Who your customer is: hospital, patient, etc.
What their pain point is
What’s the status quo, e.g. workflow for the customer
How (if at all) will the insurance cover (some of) the patient’s costs
3. Trying to integrate with EHRs.
Just because there is a ruling for EHR interoperability, this does not mean (unfortunately) that this data will be easy to access/use.
TIP: I share advice on why that’s hard and what to try instead in this blog post: “Dear Advisor: Why is it so Hard to Integrate with EHRs? (Or) Why I Don’t Recommend Integrating with an EHR, especially in your MVP”
TIP: Consider buying over building; e.g., optimize to “buy” anything that’s not core to your product, and “build” your moat.
4. For Medtech start-ups, understand the differences between submission types (pre-sub vs 510(k) vs De Novo), and risk types for Software as Medical Device (SaMD) of Class I-III, and the pathways to get the designation.
5. ML/AI model always does better if you include business context in it.
For example, if you’re developing an ML/AI model to predict time-to-ER admission, don’t forget to account for the fact that patients with abnormal heart rates or breathing patterns get priority in admission.
What are some of the most promising applications?
[Short term] improved efficiency in workflows, to help prevent/reduce burnout in clinical staff. This may seem to be the last sexy, but is one of the most pressing needs in healthcare today, with so many staff shortages.
[Medium Term] Early/better detection of conditions.
TIP: Make sure to take extra care to prevent data leakage. For example, including treatment by a physician as part of the training data (such as labs or medications), will make the (offline) model look good, but it won’t perform as well live when that information is not available since the physician will still be evaluating symptoms before assigning treatment when they’re evaluating different treatment options for a patient.
How to decide on whether to use AI/ML in the product?
To help you decide, here are some prerequisites:
Are you planning to go after a SaMD designation? If yes, keep reading; if not, skip AI/ML (at least until you’re ready to change the designation).
Are you developing an MVP? If yes, consider this opinionated blog post on why I don’t recommend AI in your MVP.
Understand your customer pain point and their workflow of how they solve it today, e.g. what’s the “status quo” now, without your product?
Does a “gold standard” exist? Will you be trying to reproduce it or define it? If you’ll be trying to reproduce it, keep reading; if not, consider the jobs-to-be-done-framework to understand better what a “(gold) standard” may need to look like.
Now that we know the end goals for the AI/algorithm, here’s advice on how to get started:
Start small! Figure out if there are any promising patterns – manually! – that may suggest/predict the desired outcome. For example, by looking at slides of biopsies of a condition, is there a heuristic/rule-of-thumb that you can use to indicate something may be going on?
Try implementing the heuristic with if/then statements. Note: This can typically get you 80% there in some situations!
Evaluate, paying special attention to false positives (predict something happened when it didn’t) and false negative risks (missing an event that happened), especially relative to a “gold standard” if it exists, and the ability to detect the difference in treatment between the outcomes! Few do this! Then, determine what needs to improve.
TIP: Don't forget to use training, validation, and test data sets to evaluate the algorithm, reporting the final performance metrics based on the never-before-seen test data.
Try a slightly more sophisticated algorithm, which accounts for (some of) what needs to improve. Evaluate!
TIP: Focus on the simplest algorithms that are easier to verify, understand, and explain. It will be easier to implement and maintain, and as a side effect, it may be easier to get FDA approval as it’s more transparent.
Advice for HealthTech/MedTech Start-ups Looking for Funding
First, we need to understand how funds work: LPs give money to funds because they expect to diversify their portfolio of investments and get better returns than the market; VCs invest when they expect a high enough ROI so that their 2 and 20 fund fee compensation structure makes sense.
TIP: Funds are looking for you to derisk your idea before investing; they’re not looking for you to de-risk your idea with the fund’s money…
Consider applying to non-dilutive accelerators and for non-dilutive grants to de-risk the idea before fundraising from VCs.
Focus on understanding your customer pain points and workflows, then lean into technology/AI to help scale that idea.
That is, don’t be that start-up trying to reduce physician burnout, doesn’t shadow them, and misses capturing the pagers they get. (True story!)
Don’t be that Biostatistician that’s helping a PI reduce the time it takes to do rounds, but one that doesn’t understand what happens in every minute of each session. (True story!)
Because… AI not delivered as a data product leads to hallucinations and nightmares. :)
Consider having a team (which includes yourself and advisors) with expertise in all of the following: clinical context, strategy, and execution.
"AI not delivered as a data product leads to hallucinations and nightmares." -- Irina Kukuyeva, PhD
You May Also Like
Talk: Power of Data for Product Market Fit (or) Your Product Market Fit Roadmap to Escalate Your Growth (including slides and recording)
(Step 1) Dear Advisor: How do I get Started on a Data Strategy from the Beginning?
Founder Institute's "Health Tech Startup Resource List: 645+ Accelerators, Incubators, Investors, and more" (2022)
FemTech Lab's webinars on:
"Breaking into the US: HealthTech: Best Practices for Founders" for (so many!) things to know, consider, and watch out for as you GTM