Dandelion Health is a technology-driven healthcare company that leverages artificial intelligence to provide innovative solutions in the medical field. The company focuses on improving patient outcomes, optimizing healthcare operations, and offering data-driven insights for medical professionals.

Dandelion Health

Dandelion is a product development platform for clinical AI, focused on harnessing the immense power of healthcare data to create innovations that will benefit all patients and communities

"Older adults, a population that will reach 80 million in the U.S. by 2040, are often underrepresented in the data used to train and test healthcare algorithms. This gap has historically exacerbated existing social inequities, especially for communities of color, low-income individuals, and those residing in underserved areas."

We are delighted to partner with The SCAN Foundation (TSF) to add social determinants of health variables to our free public service for algorithm audits. Dandelion is already auditing commercial and academic algorithms for performance and fairness by age group and self-reported sex and race/ethnicity. With TSF's support, we will soon launch the ability to assess algorithm bias across patients representing different income levels, social vulnerability index scores, and urban/rural zip codes.

For health systems thinking of implementing AI algorithms, or the developers who want to sell to them, this is a Must Do. 

TSF joins Gordon and Betty Moore Foundation to generously seed this program and push the scientific frontiers of algorithmic fairness in healthcare. 

Interested in participating or learning more? 

Read the press release: https://lnkd.in/dSrxZDx8
Visit our Validation webpage: https://lnkd.in/g99Xa486
Get in touch using our contact form: https://lnkd.in/g23jwAf5

We're proud to be named as one of Digital Health New York's 10 Companies to Watch in 2024, alongside "the most innovative and exciting health start ups in the New York area," as DHNY's announcement put it.

We believe in the potential of AI for a better patient experience and healthcare system — one that's more equitable and fair for all — and we're glad to share that commitment with our fellow honorees.

https://lnkd.in/gzvtb98A

"AI will transform medicine and the health care system— for better or for worse, depending on how it is built and applied."

Our Co-Founder and Chief Scientific Officer Ziad Obermeyer testified before the U.S. Senate Committee on Finance to discuss how the government, private sector, and researchers could advance commonsense approaches that both protect patients and promote AI innovation in healthcare. While there are several specific things that programs under this Committee’s jurisdiction could do to ensure that AI produces the social value we all want, Ziad highlighted three approaches in particular:

1️⃣ "AI developers must be transparent about the output of their algorithms. If an algorithm predicts health care costs, the developer should not be able to claim that it predicts “health risks” or “health needs”—unfortunately, many cost-predictors currently do exactly this."

2️⃣ "Algorithms’ outputs should be evaluated for accuracy in a completely independent dataset, both overall and in protected groups, in keeping with good machine learning practice [so algorithm purchasers, users, and regulators] can focus on one simple question: is the algorithm predicting what it's supposed to predict, accurately and equitably?"

3️⃣ "AI products should be valued and reimbursed according to established principles from health economics and outcomes research. If an AI results in an earlier diagnosis of heart attack or breast cancer, for example, that generates value to patients in the form of life-years, and to the health care system in the form of downstream costs avoided. The sooner public programs lay out what they are looking for, the sooner the market can deliver safe and effective AI products to solve the urgent problems they face."

The full Committee hearing is available here: https://lnkd.in/eXBGjN4q

While upstream AI-enabled drug discovery gets a lot of air time, an under-played impact of clinical AI is the extent to which it will revolutionize pharma’s use of real-world data (RWD) for downstream drug development and commercialization (patient cohorting for clinical trials, better insights for label extension or drug repurposing).

RWD has traditionally been focused on structured data (claims, ICD-10s, labs, vitals, prescriptions, etc.) or labor-intensive abstraction of unstructured data (e.g., physician notes) into structured data. LLMs are going to drop the cost of clinical note abstraction by at least an order of magnitude, opening up a whole new range of use cases for such data (e.g., looking at provider notes to identify reasons for treatment uptake, switching or discontinuation).

But it’s not just language-based modalities that will benefit from this. A computer-vision algorithm developed to automate the quantification of clinical measurements on a radiology scan can also be used to abstract structured metrics from the petabytes of imaging data generated by health systems. Instead of looking at claims data or diagnosis codes (all designed for administrative, not clinical purposes), pharma companies will instead be able to use AI to directly examine raw clinical data from imaging and waveforms at a broad scale that would be prohibitively expensive with human labor (e.g., did the treatment actually improve heart function or liver health, as seen in the ECG or MR scan?). 

At Dandelion, we have a number of projects underway in this space already, and I’m quite excited to see what they yield. This feels like a big, big opportunity to push this space forward.

Thoughts? Drop a comment or DM me so we can discuss more.

Over the past few months, I’ve had several conversations with radiology AI developers large and small who are working on foundation models. A few of the tools are incredibly technically impressive (and getting better fast), but I wanted to share a few observations about the second-order consequences of how these models are being trained and built.

1.    	These models are being built with large volumes of data, but those data are fragmented – mainly DICOM/radiology report pairs without additional clinical context. While this may enable automation use cases (“here’s a DICOM, draft the radiology report”), this business model has not proven to be a winner thus far. Until the technical, social, regulatory, and liability issues with truly autonomous clinical AI are sorted out (not holding my breath), partial automation is a feature, not a product. And good luck getting a payer to reimburse for your product feature.

2.    	To me, the much more interesting and high-impact potential use cases for clinical AI are around prediction and prognosis – actually doing things that radiologists can’t do today. Foundation models ought to be great at finding these signals! Yes, the scientific and regulatory bar is a lot higher, but the reward for doing the work is a product that actually has a viable business model – payer reimbursement. However, these use cases require longitudinal datasets that situate the images in the context of broader diagnostic / treatment pathways, and these are much harder to come by. Many of the developers I speak to have trouble even sourcing serial radiology studies, much less multimodal patient journeys with outcomes. And so the tail (data) wags the dog (the valuable use case).

3.    	Developers are using foundation models to develop diagnostic algorithms much, much faster than with supervised methods. While this is cool and useful, it moves the bottleneck to the next step in the development process, which is regulatory approval. If the FDA continues to approve algorithms one at a time for individual diagnostic tasks (vs. entire suites of tasks), foundation-model-enabled development will just create a huge backlog of applications with the FDA.

Thoughts? Leave them in the comments!

AI:DANDELIONHEALTH

Dandelion Health