Has Machine Learning Finally Provided Practical Value for the Healthcare Ecosystem?

Hersh Sagreiya, MD, is an Assistant Professor of Radiology at the Hospital of the University of Pennsylvania.

Hersh Sagreiya, MD, is an Assistant Professor of Radiology at the Hospital of the University of Pennsylvania, specializing in abdominal imaging. He is board-certified in both Diagnostic Radiology and Clinical Informatics. Prior to joining Penn, Dr. Sagreiya was a fellow at the Stanford University School of Medicine. Dr. Sagreiya graduated from Harvard College with an AB in Biochemical Sciences in 2007 and earned his MD from Stanford Medical School in 2012.

Q. In 2018, after weeks of trying to follow up on a request for a machine learning analogue that provided proof points of creating value in healthcare, a retailing case was provided to me. Although the industry was different, the case highlighted how they deployed machine learning techniques to estimate historical lost sales and predict future demand of new products.

Findings included how sales did not decrease due to implementing tool-recommended price increases and that there was an estimated increase in revenue of the test group by approximately 9.7% with an associated 90% confidence interval of (2.3%, 17.8%).¹ As many of the earlier analogues shared with me were mostly marketing fluff like “reimagining the current state processes,” the healthcare stakeholders agreed to dig deeper into this retailing case.

Fast forward to 2024 and much progress has occurred around the globe. For example, given Penn Medicine’s focus on breakthrough discoveries that enabled the development of effective mRNA vaccines against COVID-19,² proton therapy to fight cancer,³ and new treatments for Alzheimer’s disease,⁴ what have you seen in terms of innovative applications of machine learning in the healthcare ecosystem?

A. Sagreiya. As an Assistant Professor of Radiology, I have been working with a team of researchers at Penn to develop the AInSights platform that provides precision analytics for medical imaging data.⁵ This collaboration involved engaging with a multi-disciplinary team of researchers to develop an AI software platform, which can automatically quantify information from medical images (such as organ volume and the presence of certain diseases) and merge this information with clinical data from the electronic health record and genetic data from the Penn Medicine BioBank.

This platform has been deployed at Penn Medicine and performs real-time evaluation of conditions such as fatty liver disease, splenomegaly, and sarcopenia. This is an end-to-end platform, and the results are available to the radiologist to dictate into the radiology report.

We have also used this information to better understand diseases (such as the clinical entity known as lean steatosis),⁶ better understand the genetics behind diseases (such as for fatty liver disease)⁷ and predict clinical labs (hemoglobin A1C)⁸ for opportunistic screening of disease. This project won the CIO 100 Award, which recognizes 100 teams that use IT in innovative ways.⁹

The vast amount of clinical data available in the healthcare ecosystem offers major opportunities for biopharmaceutical companies. Hundreds of AI/ML-enabled devices have met the FDA’s premarket requirements, and the majority are in the field of radiology.¹⁰

Many of these devices focus on medical diagnostics; however, large datasets in healthcare containing genetic, clinical, and laboratory information can also be used for preclinical drug discovery using machine learning.¹¹ Machine learning can also be used for disease prognostication, predicting clinical outcomes for cancer or whether a treatment will be effective.¹²

In the future, machine learning will enable medical care to be personalized to the patient, finding treatments tailored towards patients’ specific genetic and clinic factors. It is estimated that primary care physicians would need over 26 hours per day to provide all guideline-recommended primary care, and machine learning could play a role in filling in these gaps.¹³ In the future, one could imagine a “clinical data scientist” providing a new role for healthcare, integrating multimodal datasets throughout the healthcare system with AI to provide optimal care for different patient populations.

Even though the readership might assume that the radiology sector has made significant strides with deploying AI/ML into its daily operations, I would point out an interesting article which referenced that while progress has been made with AI integration in radiology, “…as of 2021, only 30% of radiologists reported clinical AI use, with over 70% expressing reluctance to invest in AI.”¹⁴

So, there are plenty of future growth opportunities within this segment of the healthcare field. Moreover, given the complex legal, regulatory, and compliance requirements in healthcare, forward-thinking C-suite executives from other industries will recognize the value of studying these analogues. Perhaps today, the retail industry can take a page from our healthcare sector?

Q. As biopharma firms continue to enter into collaborative arrangements with premier entities including Penn Medicine,¹⁵ what are your top three recommendations for their executives to consider when discussing emerging technology partnerships?

A. Sagreiya. First, it is crucial for both organizations to identify the key problem that they want to solve. A technology partnership will not be successful if it does not solve a meaningful clinical problem. For instance, an artificial intelligence algorithm that detects a finding obvious to a third-year medical student will not see significant adoption.

In addition, the potential solution needs to fit into a clinical workflow or physician practice. A medical device that is cumbersome or time-consuming to use will likely be ignored. Moreover, the potential solution needs to meaningfully improve upon competitors’ products. A drug that offers marginal benefits compared to other options will struggle to achieve clinical usage.

Finally, make sure that your ideas help solve pain points that resonate with the target stakeholders.Your biopharma readers should understand the key performance indicators that physicians face. Unlike their frontline staff who might be measured on their daily interactions with doctors and other healthcare professionals, most physicians are measured on an annual RVU (relative value units) target.¹⁶

Second, any partnership must engage the key stakeholders at both institutions. At the medical institution, those stakeholders include physicians and nurses who work in the field, the division head or department chair, senior hospital leadership, and patients or patient advocates. At the biopharmaceutical organization, the target stakeholders range from subject area experts to senior management.

After engaging with these stakeholders, it is important to create a collaborative plan that outlines the key goals of the partnership, which will include a timeline with associated milestones, as well as an associated budget. After a series of revisions, this document will then go through the finance and regulatory arms of both institutions until a satisfactory agreement is created that is mutually beneficial.

Finally, collaborations are typically successful when they can leverage the key strengths between the two organizations. For instance, medical institutions will have the leading researchers in a particular clinical area, whether that involves the latest treatments for lymphoma or orthopedic surgery using the newest surgical techniques. They will also have access to patient populations who could benefit from new innovations, which is essential for running clinical trials.

On the other hand, biopharmaceutical organizations will typically have access to more resources than a typical medical institution, which will be necessary for pursuing FDA approval of promising treatments. They will also have expertise in areas that a typical academic institution cannot provide that are essential for the eventual successful clinical translation of any treatments, such as market identification, creation of a business model, and day-to-day operations.

About the Author

Michael Wong is a Part-time Lecturer for the Wharton Communication Program at the University of Pennsylvania. As an Emeritus Co-President and board member of the Harvard Business School Healthcare Alumni Association as well as a Contributing Writer for the MIT Sloan Career Development Office, Michael’s ideas have been shared in the Harvard Business Review and MIT Sloan Management Review.

References

1. Ferreira, Kris J., Lee, Bin Hong Alex, and Simchi-Levi, David, Analytics for an Online Retailer: Demand Forecasting and Price Optimization, Harvard Business School, 2016

2. https://www.nobelprize.org/prizes/medicine/2023/press-release/

3. https://www.pennmedicine.org/cancer/navigating-cancer-care/treatment-types/proton-therapy

4. https://www.pennmedicine.org/news/publications-and-special-projects/penn-medicine-magazine/fall-winter-2023/a-changing-picture-for-dementia-after-decades-of-research

5. https://pubs.rsna.org/doi/10.1148/radiol.223170

6. https://www.nature.com/articles/s41598-023-49470-x

7. https://www.nature.com/articles/s41588-022-01078-z

8. https://link.springer.com/chapter/10.1007/978-3-031-46005-0_5

9. https://event.foundryco.com/cio100-symposium-and-awards/awards/?fbclid=IwAR1VeL21WFfj0Dj-nNLAA_xIjFKF6ui2LAiaaYDequFVnob2IX4AEcrex7E

10. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices

11. https://www.nature.com/articles/s41589-024-01679-1

12. https://ascopubs.org/doi/10.1200/CCI.20.00072

13. https://link.springer.com/article/10.1007/s11606-022-07707-x

14. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10487271/

15. https://www.pennmedicine.org/about/transforming-patient-care/awards-and-accolades

16. https://www.acr.org/Member-Resources/yps/YPS-News/Appreciating-the-Complexity-of-RVUs