PALO ALTO, Calif.--(BUSINESS WIRE)--Mar 19, 2025--
Hippocratic AI, the company that pioneered the first healthcare LLM Agents for patient-facing non-diagnostic clinical tasks, today, on the company’s two-year anniversary, is announcing the launch of Polaris 3.0. The 4.2T parameter constellation of 22 LLM models has delivered the safest healthcare LLM for patient-facing clinical tasks, achieving a clinical accuracy rate of 99.38% compared to 98.75% for Polaris 2.0 and 96.79% for Polaris 1.0. Polaris 3.0 was developed using extensive feedback from actual patients and their healthcare providers. As a result, the advanced model has driven improvements in patient engagement metrics, most notably increasing satisfaction scores from 8.72/10 with Polaris 2.0 to 8.95/10 with Polaris 3.0.
“Since the founding of Hippocratic AI, we partnered with health systems and clinicians to ensure our AI agents were safe and effective enough to use in patient-facing clinical operations. After hiring 6,234 US licensed clinicians to test our product with over 307,038 test calls, and incorporating the learnings of real-world evidence from over 1.85 million patient calls made in Polaris 1.0 and 2.0, we have achieved a new milestone in safety unmatched by any other Generative AI healthcare agent,” said Munjal Shah, Co-founder and Chief Executive Officer.
Alongside this release, the company published a detailed paper outlining its Real World Evaluation of Large Language Models in Healthcare (RWE-LLM) safety framework, a novel LLM testing and roll-out methodology for healthcare. The company hopes that others can benefit from the pioneering work and make healthcare AI safer. Read more about the paper here.
There are many new or improved features in Polaris 3.0. These features resulted from real-world observations of patients interacting with our AI agents in the over 1.85 million patient calls that were completed using Polaris 1.0 and 2.0. These include:
- Deep Thinking Models: Enhanced models that triple-check labs, medications, and escalations. These new “offline” thinking capabilities make a significant contribution to removing the long tail errors occurring in prior Polaris versions.
- Improved Clinical Documentation: Models that ensure health forms, including Health Risk Assessments (HRAs) and follow-up items, are documented accurately even when patients’ inputs are unclear. For example, Polaris 2.0 had a 90.5% HRA documentation accuracy. Polaris 3.0 is 98.5%.
- Advanced Emotional Quotient: Unique features like reading between the lines, multi-call memory, or suggestions for finishing the sentence for a patient if they cannot articulate quite what they are feeling, likeability, unique patient emotional adaptation, and appropriate assertiveness all helped to lift patient’s comfort in confiding with the AI agent from 88.93% in Polaris 1.0 to 94.60% with Polaris 3.0. Average call duration increased from 5.5 minutes to 9.5 minutes with the introduction of these new features showing stronger patient engagement.
- Robust Audio Handling: Real-world phone calls have a host of issues such as television, background noise, or patients mumbling words that sometimes prevent the successful completion of clinical objectives. Polaris 3.0 shows multiple improvements to speech recognition, significantly reducing error rate, that include:
- Background noise engine to help eliminate confusion by performing speech isolation (9.3% on Polaris 2.0 to 2.3% on Polaris 3.0).
- New speech detector engine to focus on the primary speaker in loud environments, this also includes when the TV is playing loudly in the background (15.0% on Polaris 2.0 to 2.4% on Polaris 3.0).
- Novel single-word engine occurs when LLMs struggle with hard to understand single-word answers because they don’t have other contextual clues (2.4% on Polaris 2.0 to 0.2% on Polaris 3.0).
- Improved entity transcription engine for medications and numbers since accuracy here is critical (4.2% on Polaris 2.0 to 0.5% on Polaris 3.0).
- Clarification engine is a novel feature that clarifies what a patient says in a graceful and socially acceptable manner. (16.3% on Polaris 2.0 to 2.0% on Polaris 3.0).
- Multi-lingual Safety Equivalency for Spanish: The Spanish version is now at a 99.83% accuracy of giving the right answer. Overall, across nine non-English languages - Arabic, French, Hindi, Japanese, Korean, Mandarin, Portuguese, Russian, and Spanish - the accuracy is 99.09%. The company has also added novel features like multi-lingual auto switch. The feature allows the AI agent to start speaking in Spanish if the patient does, even if Spanish is not listed as the patient’s primary language.
- Orchestration Features: Besides the actual patient call, the company has added many features needed by health systems, payors, or life sciences companies to ensure Hippocratic AI agent calls integrate with clinical workflows. These include: navigating IVRs of other providers, labs, or pharmacies; accurately quoting policy documents like explanation of benefits (Polaris 2.0 is 86.4%; Polaris 3.0 is 99.4% of the time); scheduling of complex appointment scenarios (error rate of Polaris 2.0 is 8%; Polaris 3.0 is 0.5%); and handling of adverse event reporting and ensuring no conversation of off-label use for pharmaceutical clients.
- Dialer Features: Successfully connecting with patients is required to complete patient objectives. Polaris 3.0 adds the ability to leave voicemails, pause to allow patients to complete blood pressure readings, resume calls if a call is dropped, send text messages, call back at a given time, passing all context to any human we escalate the call to (ANI), and making warm and cold call transfers.
- Deeper Integrations with EMRs: Polaris 3.0 has now been successfully integrated with health care systems of records such as Epic, Cerner, and Salesforce with the ability to integrate with other major and specialty systems Athenahealth, eClinicalWorks, Nextgen, Modernizing Medicine, Allscripts, Meditech and more.
These new features have lifted key patient engagement statistics to new highs in Polaris 3.0:
- Connect Rate: New high water mark of 89.28% with Polaris 3.0.
- Call Completion Rate: New high of 96.46% with Polaris 3.0.
- Refusal to Talk to AI: Refusal rate decreased further to an average of 2.68% with Polaris 3.0 from 6.06% with Polaris 2.0.
- Patient Satisfaction: Both versions have achieved high patient satisfaction ratings, however, Polaris 3.0 had a higher average score of 8.95/10 compared to 8.72/10 for Polaris 2.0.
“Vertical AI agents require features that are unique to that specific environment and handle the long tail of issues. Our goal for Polaris is a level of product perfection to ensure that our products meet or exceed the rigorous requirements of real-world clinical and patient environments, and are not just a novel AI tool.” said Subho Mukherjee, Co-founder and Chief Scientist of Hippocratic AI. “While Polaris 3.0 release gets us much closer to that goal, it is one we will continue to relentlessly pursue.”
Polaris 3.0 will be made available April 7.
About Hippocratic AI
Hippocratic AI has developed a safety-focused Large Language Model (LLM) for healthcare. The company believes that a safe LLM can dramatically improve healthcare accessibility and health outcomes in the world by bringing deep healthcare expertise to every human. No other technology has the potential to have this level of global impact on health. The company was co-founded by CEO Munjal Shah, alongside a group of physicians, hospital administrators, healthcare professionals, and artificial intelligence researchers from El Camino Health, Johns Hopkins, Stanford, Microsoft, Google, and NVIDIA. Hippocratic AI has received a total of $278 million in funding and is backed by leading investors, including Andreessen Horowitz, General Catalyst, Kleiner Perkins, NVIDIA’s NVentures, Premji Invest, SV Angel, and six health systems. For more information on Hippocratic AI, https://www.hippocraticai.com
View source version on businesswire.com:https://www.businesswire.com/news/home/20250319172281/en/
CONTACT: Media Contact
Rick Keating
917.767.2400
rkeating@keatingco.com
KEYWORD: UNITED STATES NORTH AMERICA CALIFORNIA
