The State of AI in Healthcare PART I
With the ever-growing buzz surrounding AI, startups, and investments, it is crucial to carefully consider its implementation in the healthcare industry. While AI has been utilized in healthcare for some time now, with healthtech companies leveraging machine learning algorithms and automation, it has come a long way in the past decade. Particularly in the last five years, AI in healthcare has become highly specialized, catering to the demanding and rapidly transforming US healthcare system.
Despite the economic challenges, healthcare and health-tech remain one of the strongest industries, if not the strongest, in the market. Although health-tech is not immune to fluctuations in valuations, investments in the range of 15-25 billion dollars are expected to continue in 2023, similar to the levels seen between 2020 and 2022. Furthermore, AI is understood to make a bigger impact in healthcare than a traditional tech software businesses because margins are low to begin with for healthcare services compared.
In this article, we will explore some of the various healthcare segments (I’ll definitely need more parts to this), albeit not an exhaustive categorization, and how AI is shaping them through exciting new projects.
SAMD- Software as a Medical Device
AI in healthcare has progressed beyond image detection, now encompassing sound as well. Sound serves as a precursor to language, which plays a significant role in the rise of Language Models (LLMs), more on this below.
Novoic, in partnership with leading institutions, is making early testing for Alzheimer's disease accessible and scalable. Leveraging software and natural language processing, Novoic offers a user-friendly audio-verbal assessment on any smart device. The AI behind Novoic's technology is backed by extensive research in the field. Its results are comparable to plasma tests, which are more challenging to administer and time-consuming. Additionally, Novoic has made significant progress in collecting relevant and diverse data to enhance its algorithms. It provides a comprehensive software platform to monitor patients at scale.
The primary obstacle hindering the validation, success, and further development of SAMD is the scarcity of testing and evidence in the market. Many SAMD companies rely on the potential success of future clinical trials and partnerships with health institutes. These trials in the next few years will further validate the AI in one way or another and largely determine how this segment moves forward. Moreover, the participation of payors and the integration of these treatments into health plans are still uncertain factors, posing significant challenges.
Administration- Mental Health Care
The field of mental health has experienced significant growth in recent years across all aspects of the care pipeline. Efforts have primarily focused on bridging the gap between patients and therapists/care providers. Lyra Health, Spring Health and Better Up are prominent companies in this space, working to shorten this bridge. Additionally, platforms like Headspace, Calm, and Balance have focused on making mindfulness and meditation more accessible as part of the treatment process. While there is some AI involved on the backend of these platforms, such as matching algorithms and recommendation engines, there has been limited progress outside of traditional methods for diagnosing mental health issues. This is where AI Berry comes into play.
AI Berry is a platform to make screen for mental health disorders more accesible. Combining predictions based on speech, facial expressions and text inputs, AI berry uses a fusion machine learning model design and assists clinicians and providers in making a diagnosis and assigns risk levels to each patient. These models make it easier to bring all the different inputs in the screening process together in one format and produce a single output such as a depression risk score. In addition to assitance, AI berry is also providing a self-assesment for individuals. To assess the reliability of the platform, AI berry has partnered with various institutions to validate the platform with clinician based results. By pooling inputs from multiple institutions, the platform has a real chance at building a robust and accurate tool for mental health diagnoses.
Large Language Models (LLMs)
Perhaps the most definitive and low hanging fruit of the LLM use in healthcare will be on the healthcare administration and assitance side of things rather than the more direct clinical applications. There are several companies that are already utilizing NLP and Machine learning and have made significant progress in a short time.
Leading the way is ScienceIO. Science IO is something that any person even remotely involved in the healthcare system needs and this includes PMs like me working in HealthTech. Having experienced the pain of going through Medicare documentation that can span 100s of pages to come up with half a page of summary, I have experienced this pain first hand too many times. ScienceIO’s platform, using billions of data inputs and a simple input (text) to output (JSON) model, cateogrizes more than 9 million medical terms out of almost any unstructured text (including EHR, Clinical Trial Documentation) form and is easily accessible by a scalable API. One of the most valuable use cases that Science IO is able to solve is managing PHI. Any project involving assessing or cleaning up a new type of dataset can utlize Science IOs platform to extract and redact PHI.
A big challenge also that I have experienced on the Real World Data/Evidence side is the lack of cleaner and accurate unstructured data. While unstructured and “accurate” might sound conflicting, there is room for improvement on the generation of this data. No physician and health clinic is thinking about generating better EHR data to help a RWE company commercially improve patient care, there are several pain points earlier in the funnel itself. One of the main ones is accurate information sharing for physicians and patients and the transmission of that information between them. A company that has made big strides toward this problem is Abridge. Abridge is using science backed a Audio Speech Recognition system (ASR) (just take a look at their research contributions) to record patient doctor conversations. ASR’s are everywhere from Siri to cloud offerings by Google and Amazon, but Abridge’s system uses a combination of text and speech fusion to produce better results in a healthcare specific setting. In addition to the audio recorded from the conversation, Abridge leverages a healthcare library for the system to constantly look for errors. Error reduction is one of main ways ASRs are being evaluated and Abridge has achieved improvements upon a strong baseline by up to 15.4%. With the improved ASR, doctors have reported having better notes and patients have a better understanding the medicine heavy language that often plagues them during their care.
Another problem Abridge is helping with to some extent is reducing physician burnout. Burnout peaked after pandemic, with 63% of physicians reporting some sort level of burnout. Burnout in Healthcare is also speciality based with some physicians experiencing it more seriously than others. One of these specialties is Radiology, in which doctors spend a significant amount of time reporting their findings and recommendations. Rad AI reads the core findings of a radiologist during dictation and composes a unique impression that belongs to the particular radiologist. In the report, it pulls in all significant findings, highlights pertinent negatives, and answers the main clinical question. This significantly helps with fatigue and improves accuracy of reports in these settings.
Summary
AI is far from a novel tool in Healthcare or Digital Health. Over the past decade, it has played a significant role in data-driven companies. Nonetheless, the emergence of Generative AI and the emphasis on LLMs will undoubtedly amplify the focus on AI in Digital Health. The forthcoming years will prove crucial in determining AI's triumph in these sectors and how innovators prioritize its prominence.
There are certainly other segments in health-tech such as Real World Evidence, Insurance Payor products as well as other forms of administration (outside of Mental Health) which I plan to discuss in my next post. The use of AI in some of these segments is more nuanced and serves as a supporting technology rather than the central part of the product itself.