Re-Imagining the Healthcare Delivery Journey with Generative AI
Intro
So much progress has been made since we shared our first article on the state of generative AI in healthcare.
It is amazing to see how much of our previously shared 2x2 is already green within 3 quarters. Green to us represents not just venture dollars, but customer dollars flowing into AI products in the space. We are encouraged by the potential leapfrog technologies that have been built, and, perhaps uniquely, how many of those solutions have been on the provider side, a segment semi-notorious for long sales cycles, endless pilots, and competing stakeholder paralysis. Our current read is that these AI solutions have risen to the top of the prioritization stack because they are solving need-to-have problems and providing ROIs that are measured in months, not years.
On the other hand, much of the 2x2 remains yellow and gray. This is not to say that innovation is not happening in these areas; we have met with fantastic entrepreneurs building companies in each of these segments and expect more entrants in these areas. While certainly possible we have missed a breakout company or two in these areas (if so, please reach out!), our observations are that the customer traction is trailing that of the green segments. And in the gray areas, our view is that the best innovation is yet to come.
Diving Deep into the Patient Journey from a Provider’s Perspective
Given the early embrace of AI-driven solutions by providers, we have decided to dive deeper into this area.
Leveraging our diverse expertise in healthcare, AI, startups, and investing, our team –which includes a practicing physician – has set out to break down the patient journey from a provider’s perspective to get more granular in terms of ways that AI could transform patient care.
We hope that by delving into this level of detail we can equip founders with information to spark meaningful advances in healthcare delivery.
In this report, we:
Share detailed process maps of the healthcare delivery journey
Highlight steps at which companies are already innovating and others where we think there is substantial whitespace.
Discuss the LLM use cases we find most exciting and those that may be best addressed by startups
We hope that sharing this level of granularity is helpful to both seasoned health tech veterans and to those with less healthcare experience hoping to apply their tech expertise to improving our healthcare system.
Structure
We have organized our care delivery map into 3 sections according to the typical patient journey:
Pre-hospital care
Inpatient (intra-hospital) care
Post-hospital care
We have further subdivided each main section into subsections as can be seen in the below graphic.
We will share a detailed healthcare delivery process map for each subsection and include analysis on both the steps at which AI companies currently operate and those steps at which we see lesser explored opportunities.
Pre-Hospital Care
Outpatient Care
In this article, we will focus on outpatient care, and, in our market map below, we highlight some of the companies we discuss. Future posts will include market maps for the remaining sites of healthcare delivery (e.g. ED, inpatient, post-hospital) and deep dives for their respective companies.
We’ve also included a table below that contains the highlights for the remainder of the article. For two similar summary figures about the question of who wins, incumbents or startups, for which use cases, please see our summary figures section at the end of the article.
Pre-Visit Planning
Legend:
Appt. = Appointment
RN = Registered Nurse
EMR = Electronic Medical Record
UC = Urgent Care
ED = Emergency Department
PCP = Primary Care ProviderA Walk Through the Process Map: Pain Points, Current Solutions, and LLM Use Cases
Patients enter the outpatient healthcare system mainly through two paths: 1) health system-initiated follow-ups or 2) patient-initiated, symptom-triggered calls. Both routes face challenges like appointment scheduling, patient intake, triage, and the need for physicians to sift through extensive patient data. Despite several startups working on these issues, significant opportunities remain for generative AI to make further improvements.
Health System-Initiated Scheduling
As per the process map, navigating the outpatient healthcare system often starts with a health system-initiated follow-up appointment, a process historically fraught with scheduling and intake challenges. Companies like Zocdoc, PatientPop (acq. by Kareo), and Phreesia have made strides in connecting patients with available slots and streamlining intake, yet the problem persists.
Enter AI: solutions like Artera’s patient communication platform and Fabric’s care navigation assistant are at the forefront, using AI to enhance follow-up adherence and patient engagement.
“In healthcare, an industry beset by data isolation … AI-ingestion systems can harmonize disparate datasets – from consumer demand and real-time capacity to seasonality – to make data-driven decisions about how resources are allocated on a given day, month, or quarter. The result? More access to care, better treatment, and the precise modulation of capacity.” —Derek Streat, CEO of DexCare
Despite efforts from many startups to tackle scheduling and patient intake challenges, these issues persist in healthcare. A 2022 survey by Notable revealed that 61% of patients avoided doctor visits due to the cumbersome scheduling process, and 41% changed providers over poor digital experiences.
These persistent challenges around scheduling and intake, combined with AI's potential, present an opportunity for entrepreneurs. While more clinically-oriented tasks (e.g. clinical decision support or treatment recommendations), face regulatory hurdles, many non-clinical but high-hassle tasks, such as scheduling and intake, are ripe for innovation with minimal regulatory constraints.
Care Navigation Agent
We envision personalized AI care navigators as one possible solution to handle many of the frictions patients experience in healthcare. They would consist of a HIPAA-compliant app that connects to a patient’s insurance information and healthcare data (e.g. Apple Health).
These navigators could simplify tasks like booking in-network specialist appointments—either online or by phone—and automatically handle patient intake paperwork. They would arrange lab work and imaging, schedule procedures, pay bills, and ensure coordination of medication refills between your doctor’s office and the pharmacy, all while optimizing costs e.g. by efficiently navigating the use of HSA/FSA funds, cash payments, GoodRx discounts, employee benefits, and the timing around deductibles.
Later versions could summarize your healthcare visits for the year and help you arrange all the necessary appointments, testing, and lifestyle changes suggested by your doctors.
Patient-Initiated Pathway
We’ll now discuss the bottom pathway of how patients first access healthcare i.e. symptom-triggered patient calls.
Companies like Buoy Health, Ubie, Corti, and Ada Health deploy AI chatbots to more rapidly and accurately triage patients to the appropriate level of care (e.g., primary care, urgent care, ER) while reducing costs for payors. Similarly, Kyruus and Dexcare use AI to connect patients to the right providers while balancing demand and health system capacity.
However, deploying AI in patient triage presented challenges, especially in balancing the demands of paying customers i.e. large institutional clients (e.g. payors and employers) and end users (i.e. patients). Needing to build both a strong enterprise solution and a strong patient-facing solution was a significant challenge. While many startups did a great job scaling quickly with paying institutional players, they often struggled to maintain long-term engagement with patients resulting in significant churn. They realized that they needed to completely refocus from enterprise clients to finding ways to build more meaningful relationships on the patient side. There were 3 major contributors to this struggle:
It was difficult to break a patient’s habit of searching the internet for symptoms.
It was hard to engage people on a continuous basis once they’re feeling better.
It was challenging to forge meaningful, long-term patient relationships when payors and employers' established benefit access methods relegated AI chatbot triage companies to mere links on the company’s intranet.
Patient-Centered Healthcare Triage
Despite these challenges, patient-centered triage is still valuable, since we’ve seen triage startups scale quickly with paying enterprise customers, and we think LLM’s could help address their patient-facing problems - high churn and decreased engagement.
If a startup could offer additional, LLM-enabled ongoing services beyond triage, they might be able to improve retention of patients. For example, imagine adding an LLM-enabled chat to interact with your own health record or a care navigation agent to book and manage appointments. Patients will continue to seek out care when experiencing significant symptoms, the opportunity is to keep them around after those symptoms resolve.
Synthesizing Information for Physicians
Physicians are increasingly overwhelmed by the sheer volume of data in EMRs, which only continues to grow. Fortunately, one of generative AI’s key strengths is it’s ability to synthesize information.
Research at Stanford supports this healthcare use case (post, website). First, the researchers found that clinicians spend nearly half their day using EMRs. They then wanted to figure out what clinicians actually wanted from an LLM that could interact with EMRs, so they asked them to submit instructions for the LLM. Greater than 66% of clinician-shared instructions were to "retrieve & summarize" EMR data, suggesting that this is one of the use cases clinicians are most interested in despite more focus often placed on clinical decision support and care planning.
Integrating LLMs with EMRs to rapidly synthesize patient histories would offer healthcare providers a quick, comprehensive snapshot of each patient that is valuable in any phase of care (clinics, ER, hospital), significantly saving their time and potentially enhancing patient outcomes. This efficiency translates to cost savings that health systems value highly and already pay for. For example, Imprivata’s OneSign, an ID card scanner for quick computer logins, exemplifies the financial benefits of such time-saving technologies. Despite only saving 5 - 20 seconds per login, research shows significant aggregate savings: thousands of hours of physician time and millions in costs annually. Imagine how much more time could be saved with LLM-enabled patient summaries.
LLMs are more than ready. Recent Stanford research has already demonstrated the capacity for LLMs to produce patient summaries that were judged to be preferable to human summaries in terms of completeness and correctness.
Despite the edge that giants like Epic and Apple have with ready access to patient data, there may still be room for innovative startups to challenge them. For instance, although Epic is working on similar features, startups like Pieces and Navina are racing ahead to use AI to craft precise, actionable patient summaries from EMR data. We discuss the recipe for a winning summarization company below.
Conclusion
In summary, we see an opportunity for LLMs to make a significant impact in several areas with some being more easily addressed by startups and others by incumbents:
Startup Opportunities:
Care navigation agent: an LLM-based assistant handles the time-consuming frictions of a patient’s healthcare journey allowing them to focus on their health. Use cases include scheduling in-network appointments, testing, and procedures recommended by a patient’s physicians, filling out patient intake forms, managing medical expenses e.g. paying bills, making optimal use of HSA / FSA / cash pay options like GoodRx, navigating deductibles, and navigating employer benefits.
Why a Startup Opportunity? Policy changes improving digital access to patient data and LLM technological capabilities allow startups to compete in this space.
Toss-Up Opportunities (Startup or Incumbent):
Patient-centered chatbot triage: an AI chatbot triage solution that better solves the end user (patient) side with a combination of better technology and long-term, follow-up offerings beyond triage to maintain engagement.
Why a Toss-Up Opportunity? Startups have been working in this space for many years, and some have seen significant challenges or gone out of business. A new technical stack leveraging LLMs, whether from a new or an existing startup, may improve engagement and retention of patients.
Incumbent Opportunities:
Patient history summarization tools: will smoothly integrate with multiple EMRs, providing comprehensive summaries from multimodal data alongside actionable recommendations based on evidence-based medical guidelines. They will fit seamlessly into provider workflows while offering significant value to both providers and payers.
Why an Incumbent Opportunity? Hard to overcome the gatekeeping capability of large EMRs (duopoly between Epic and Cerner) who control access to patient information and physician workflows.
Intra-Visit Workflows
Legend:
MA = Medical Assistant
PCP = Primary Care ProviderA Walk Through the Process Map: Pain Points, Current Solutions, and LLM Use Cases
No-shows
The first step for the patient is arriving at the clinic. Unfortunately, before the pandemic, medical practices experienced a 40% increase in the median patient no-show rate, rising from 5% in 2018 to 7% in 2019, with about half of medical practices reporting a continued increase in no-shows between 2021-2022 before finally holding steady in 2023.
Estimates of the cost to the medical system come to $150B each year.
There are many contributors to no shows including difficulty securing time off work or finding affordable childcare or transportation, often disproportionately affecting lower-income patients and widening healthcare disparities. Patients facing challenges to their social determinants of health (SDoH) often have higher no show rates.
Companies like Ride Health, Uber Health, and DocGo are working to address medical transportation, while startups like Healthify (acquired by WellSky) focused on other SDoH. Healthify built a platform to address SDoH by tracking a patient’s social needs and coordinating referrals with community partners to improve their health, enabling payers and providers to better coordinate and manage the social needs of their patients.
Social Determinants of Health LLM Agent
While these have been moves in the right direction, there remains much room for progress. Generative AI allows for an unprecedented level of personalization and engagement for each patient. A generative AI agent personalized to a high-risk patient's health data and social situation could help them navigate resources and our complex healthcare system to better address their social determinants of health and access to care to improve their health.
Other companies using AI-automated reminders and care navigation such as Artera and Gyant were discussed in the pre-visit planning section.
Patient Intake
Traditionally, there are quite a few intake steps after the patient arrives at the clinic including: check-in, insurance verification / copay, questionnaires, and, later, check-out and scheduling of follow-up appointments.
Several companies such as Notable use AI and RPA to automate these steps, improve operational efficiency and improve follow-up rates.
“My framework for healthcare automation has always been twofold: knowing what to do and then how to do it. The what is the intelligence of knowing which sequence of actions to take and the how is the integration to perform those actions in an automated way. I think a lot of people focus on the intelligence side of things, for example, predicting which claims are at highest risk of denial, while underestimating the difficulty of the actual integration. People often just assume they can use FHIR or HL7, and that's the extent of their diligence, but it turns out true robust integration is much more difficult.” — Muthu Alagappan, MD, former CMO, Notable
While Notable offers a platform solution to automate many steps of the healthcare delivery process, other companies are more focused on automating specific steps. For example, Visit Pay (acq. by R1 RCM) focuses on simplifying the patient billing experience, while Cedar focuses on automating the pre-visit experience in addition to medical billing including appointment reminders, digital registration forms, and the ability to collect insurance information. They have also recently developed a generative AI tool to assist patients understand and resolve their healthcare bills.
Patient Care
After a medical assistant has readied the patient, the provider begins patient care. In current practice, many EMRs already feature clinical decision support (CDS) prompts that pop up to suggest actions to providers. Most of these prompts are quickly overridden and not acted upon by providers for several reasons:
Many are deemed irrelevant to the current patient
Alert fatigue from so many irrelevant alerts
Many don’t offer sufficient explainability as to why the prompt is being suggested for that specific patient and rarely offer references in a helpful manner
Automated, Real-time Relevant Medical Information
LLMs offer information retrieval, summarization, and synthesis capabilities that allow for a new-generation of workflow-integrated CDS.
LLMs could bring relevant medical information to the physician in real-time based on the current patient improving outcomes and expediting care, acting essentially as an integrated, next-generation version of UpToDate (the most used online clinical reference).
Finally, while explainability is not inherently a strong suit of LLMs, they could still be used to clarify, retrieve, and summarize information about studies that have generated the CDS information when clinicians do have questions.
Documentation
Next in our process map is documentation, a major time sink and pain for clinicians — and a solution that has been dreamed up for decades is ambient note generation. Today this technology is the fastest growing, most exciting application of GenAI in healthcare and has interest from incumbents and startups alike. Companies like Nuance (partnered with Microsoft/Epic/OpenAI now), Augmedix, Suki have been around for years chipping away at this problem, while numerous new companies, Abridge, Deepscribe, Ambience, Nabla, Knowtext, and Eleos Health have rushed into the space.
Translation of Medical Handouts
One thing that is being asked of ambient documentation companies is working in multiple-languages. Finally, a last step of a patient visit is often printing out handouts for patients. Sometimes, translation is necessary for handouts of medical information - this is an area both startups and incumbents are looking to make far easier.
Real-Time Coding and Billing
Throughout this care process, documentation is being written and orders are being entered. Startups such as SmarterDx, Regard, Nym, Nuance, Codametrix, Fathom, and Decoda Health have already done a good job of moving towards real-time coding from that entered information. The next step is real-time billing. There has been much excitement about using LLMs to parse this information to capture the encounter and submit related billing in real-time.
A major question is who is best positioned to create the winning real-time billing solution. Will it be notetaking startups moving into the billing space (given their capability to interpret notes), will it be companies with RCM expertise like AthenaHealth moving upstream to capture this information from notes, or perhaps major payors like United given their dual role in processing claims and managing healthcare services.
Conclusion
In summary, we see an opportunity for LLMs to make major impacts in several areas of intra-visit workflows with some being more easily addressed by startups and others by incumbents:
Startup Opportunities:
N/A
Toss-Up Opportunities (Startup or Incumbent):
Automated, real-time relevant medical information: LLMs could bring relevant medical information to the physician in real-time based on the current encounter improving outcomes and expediting care, acting essentially as an integrated, next-generation version of UpToDate (the most used online clinical reference).
Why a Toss-Up Opportunity? Ambient note generation companies and EMRs have an advantage given their access to notes and integration into workflows; however, being able to provide real-time, personalized, accurate and explainable medical information in a way that physicians like will be no mean feat and may require a new entrant to tackle.
Incumbent Opportunities:
Real-time billing: using LLMs to capture information from notes and orders within the patient visit to submit related billing in real-time.
Why an Incumbent Opportunity? Real-time billing presents greater challenges than real-time coding, favoring incumbents who already navigate existing billing workflows such as those with RCM expertise like AthenaHealth or major payors like UnitedHealthcare given their control over claims data.
Social determinants of health LLM agent: personalized to each patient's health data and social situation to help high-risk patients navigate resources to address their social determinants of health and thereby improve their health.
Why an Incumbent Opportunity? More of a feature that would need to be part of a more comprehensive care management solution to truly move the needle on outcomes and cost savings.
Translation of Medical Handouts given to patients.
Why an Incumbent Opportunity? More of a feature that would integrate into existing workflows in the EMR.
Referrals
Legend:
PCP = Primary Care Provider
Appt. = AppointmentA Walk Through the Process Map: Pain Points, Current Solutions, and LLM Use Cases
Referrals: What Could Go Wrong?
Specialist referrals are a critical part of the outpatient care pathway; nearly 100M are placed each year, but up to 50% are never completed (meaning the patient does not end up seeing the specialist).
Reasons for incomplete referrals include missing information, misdirected referrals, or faulty communication. For example, in one study, ~70% of PCPs reported sending patient information to specialists “always” or “most of the time”, while < 35% of specialists reported receiving that information.
This can lead to missed or delayed diagnosis and treatment, malpractice, and significant losses in revenue for specialists. For instance, between 20-30% of diagnostic errors may be attributable to breakdowns in the referral process. From a value-based care perspective, unnecessary referrals by value-based PCPs can also lead to unnecessarily high expenditures.
The graphic below by the Institute for Healthcare Improvement illustrates many of these points. Specifically, for those ideating in the space, the bottom of the graphic contains 9 steps in the referral process that could serve as a springboard for innovation.
Current Solutions and Future Directions
We see two current approaches to this problem:
E-consult providers
Referral management solutions
E-Consults
Health systems including Mass General Brigham and Kaiser are leveraging electronic consults (e-consults) to streamline care. This process allows PCPs to query specialists via EMR, receiving advice within days to implement or, if needed, to arrange a specialist visit. E-consults save time, cut costs, and enhance PCP decision-making.
Startups like Sitka aimed to revolutionize the e-consult model, offering asynchronous video and text consultations across various care settings. Their platform enabled quick specialist feedback within five hours, slashing formal referrals by 87% and delivering major cost savings in value-based care.
On the other hand, PicassoMD innovates with real-time consults integrated into PCP visits, diverging from the asynchronous approach. This model helps PCPs make immediate care decisions and enhances patient satisfaction by resolving issues in a single visit, reducing missed care due to missed follow-up appointments.
Specialty-specific LLMs (e.g. focused on specific specialties like cardiology or dermatology) may enable the next generation of a Sitka or PicassoMD-like company. These advanced LLMs promise to cut specialty costs by providing instant answers to PCPs, initially reviewed by specialists but eventually autonomously for simple, guideline-based answers. Developing these specialty-specific LLMs might involve fine-tuning a general LLM with general medical knowledge and then partnering with an academic to further fine-tune it on the gold-standard text for a particularly specialty (e.g. Bolognia’s Dermatology) as well as the consensus care guidelines for that specialty and other relevant literature.
However, the implementation of such LLMs must be carefully balanced against the preferences of the PCPs using them. For example:
“LLMs could really help, especially since 67% of the time our specialists are making medication recommendations. But to counter that point, we continuously hear that PCPs love actually seeing the video from the specialist and knowing that someone actually listened to the nuances of their patient's care.” — Kelsey Mellard, CEO, Sitka
Interestingly, it would not be too difficult to generate videos of specialists sharing their recommendations. This would involve an end-to-end LLM pipeline that would only require brief specialist review. LLM triages messages, drafts responses, a specialist reviews, and a generative model then creates a synthetic video of the specialist delivering this recommendation. Although this may be technically feasible, again it depends on the users’ receptiveness to AI.
Referral Management Solutions
Several startups, such as Par8o, ReferWell, AristaMD, and ReferralMD, are developing referral management solutions to help primary care practices find the best and most cost-effective specialists, share medical records, track appointment scheduling, and streamline the referrals process. Other companies like AristaMD offer both e-consults and referral management solutions. Intriguingly, Sitka and AristaMD recently merged to capture integrative value by combining AristaMD’s referral management solutions with Sitka’s extensive specialist network.
Conclusion
Technology is reshaping the referral process, enhancing e-consults and referral management solutions. The merger between Sitka and AristaMD highlights the potential integrative value of offering both of these services within the same company. Given that the majority of specialist questions are repetitive with straightforward guideline-based answers—like 67% of Sitka's consultations involving medication advice—LLMs hold the potential to radically transform the referral system.
This would let patients get much faster answers to their questions while making PCPs lives much easier and enable significant cost savings for value-based care providers. Champions of this model would likely include the value-based primary care physicians and payers benefiting from time and cost savings and the patients demanding much faster access to specialty care.
In summary, we see an opportunity for LLMs to make a major impact on referrals with some areas being more easily addressed by startups and others by incumbents:
Startup Opportunities:
N/A
Toss-Up Opportunity (Startup or Incumbent):
Comprehensive referrals solution: composed of specialty-specific e-consult LLMs + LLM-augmented referral management solutions. Given market demands demonstrated by the ArristaMD + Sitka merger, offering a comprehensive solution may be the winning strategy. A startup would likely begin with the e-consult offering while working to build out their referral network and referrals management solutions.
Why a Toss-Up Opportunity? Startups are well-positioned to tackle the specialty-specific LLM e-consult component, since these LLMs still need to be created. Additionally, partnerships with key academic writers of physician-trusted, currently used, gold standard texts plus rapid execution could be a significant moat to competition. On the other hand, incumbents with existing, strong referral networks may have a headstart on the referral workflow solutions component.
Incumbent Opportunities:
N/A
Patient Panel Management
Legend:
PCP = Primary Care Provider
RN = Registered Nurse
MA = Medical Assistant
PharmD = Doctor of Pharmacy
SW = Social WorkerA Walk Through the Process Map: Pain Points, Current Solutions, and LLM Use Cases
This discussion on patient panel management targets pre-hospital care, yet there's a natural overlap with post-hospital care due to the fine gray line that separates what could be categorized as pre vs post-hospital care, reflecting the continuum of patient care. Aspects of care that may seem to be missing here can be found under the post-hospital care section.
Empanelment
Patient panel management begins with empanelment, assigning patients to specific PCPs and their support teams — including MAs, RNs, PharmDs, population health specialists, case managers, and SWs. This step establishes clear responsibilities for each patient’s care, fostering accountability and facilitating proactive panel management to improve preventive care and chronic disease management.
Incentives According to Reimbursement Model
It’s important to stress the different incentives according to the primary reimbursement type at each medical practice. This significantly impacts product-market fit for patient panel management solutions.
Fee-for-Service (FFS) Model: Providers are paid per service performed irrespective of outcomes, encouraging a higher volume of tests, referrals, and procedures. Although FFS codes exist for chronic care and remote monitoring, proactive panel management is often less rigorous, since these codes only partially align incentives compared to value-based care models.
Value-Based Care (VBC) Model: Compensation is linked to patient outcomes and quality of care, promoting proactive and preventive care, effective chronic condition management, and reduced unnecessary interventions. This model incentivizes the inclusion of additional care team members to support PCPs in helping manage their patients’ health.
Ideally, patient panel management involves a collaborative team effort, with team members practicing at the top of their skill sets to boost efficiency and mitigate burnout. Unfortunately, resource-constrained practices may struggle to expand their care teams, often overburdening PCPs with responsibilities that could be better handled by lower-cost team members.
Some reasons for this include (all addressable in part by AI):
Can’t afford sufficient support staff
Not worth it given opportunity cost
Can afford, worth the opportunity cost, but unable to hire sufficient staff
Provider and Patient-Initiated Outreach
Following further along the process map, there are 2 primary ways that ongoing panel management occurs:
Via provider-initiated outreach
Via patient-initiated outreach
Regarding provider-initiated outreach, startups like ClosedLoop are providing predictive modeling and analytics to predict which patients need proactive care and engagement.
Regarding, patient-initiated outreach, patient messages would ideally be triaged to the most appropriate care team member. Unfortunately, even in well-resourced settings, this often does not occur, and most messages inefficiently reach the PCP without any filtering.
Memora Health helps address both patient- and provider-initiated outreach via its Intelligent Care Enablement Platform, which includes: proactive AI-powered care navigation via text, clinical workflow automation assisting with tasks that don’t require top-of-license clinical expertise, and evidence-based care programs designed to extend the capabilities of care teams.
A major contributor to workload and burnout in patient panel management is the sheer volume of calls among stakeholders. AI is stepping in to help. Startups like Infinitus automate calls across the healthcare spectrum including pharmacies, PBMs, hospitals, and payors. Meanwhile, Assort Health is pioneering generative AI call centers to streamline appointment scheduling.
However, neither are helping automate the communication of medical information with patients. We're excited by the possibility of developing an LLM-enabled system to fill this gap — automating both patient- and provider initiated outreach, essentially automating most of patient panel management. By integrating traditional data analytics and AI with LLMs, this system could proactively identify and then engage patients needing care, offering guideline-based advice automatically (like adjusting diabetes medication) and more nuanced recommendations after a quick provider review. Such an approach could revolutionize patient panel management.
Patient Panel Management Care Delivery Steps
On the right-hand side of the process map, we detail some of the specific care delivery steps in panel management, alongside the care team members typically involved. Unfortunately, this often does not represent ideal utilization in that there are many steps in which less specialized members could offload care from their more highly trained counterparts.
Asynchronous Care
Companies offering asynchronous care such as Curai help address many patient-initiated concerns and allow treatment of low-acuity conditions virtually, without a real-time interaction.
LLMs are well-suited for this, since such conditions often involve routine, guideline-based care. One possible solution would involve fine-tuning an LLM on a body of medical information including clinical care guidelines. This LLM could then synthesize a patient's medical history, ask them relevant questions, and then offer a personalized plan of care, initially reviewed briefly by a provider but eventually offered directly to patients for simple, guideline-based concerns.
Pre-Op / Post-Op Care
Pre-op and post-op care is another area where we see AI being developed to improve patient care. Ufonia and Hippocratic AI are working to use AI to assist with algorithmic pre-op and post-op patient calls and ongoing follow-up. This is an exciting area but is still very nascent and will need proper guardrails to ensure safety.
Medication Adherence
Medication adherence is a critical challenge and a longstanding focus for digital health innovation, involving numerous companies, including some previously mentioned. Notifications have been used for medication reminders for years — generative AI offers the opportunity for much more personalized messaging to patients which may boost engagement and compliance.
Population Health Management Solutions
While there are large incumbents providing population health management software solutions like Optum, Epic, and AthenaHealth, there are also startups in the space including those for all patient populations like Innovaccer and Altruista Health (acquired byHealthEdge 2020) and those focused on specific patient populations e.g. ConcertoCare (seniors and other adults with complex conditions) and Canopy (oncology).
Regarding population health analytics, the Johns Hopkins Adjusted Clinical Groups (ACG) System has been the gold standard for three decades. By identifying risk and tracking patients over time, it allows for proactive intervention to reduce health care costs; however, several startups have developed their own models to improve upon the ACG System.
“The Hopkins’ ACG score and a lot of other commercial algorithms are really good at predicting utilization a year out. But that still leaves a gap in predicting which patients are most likely to get sick 30 and 60 days out, which is what we want to know in order to allocate our resources to most effectively prevent exacerbations. So we've developed our own predictive algorithms that we’ve validated to be stronger than ACG in predicting short term utilization, meaning 30 and 60 day inpatient utilization, ED utilization, and cost of care. This is really important because it shows us who is at immediate risk, so that we can figure out what their problem is and how best to address it.“ — Amy Flaster, MD MBA, CMO, ConcertoCare
Patient Reported Outcomes
Tracking patient-reported outcomes (PROs) is becoming increasingly valuable in patient panel management. PatientIQ automates PROs collection and then provides actionable insights for quality improvement, research, and marketing. Personalized, conversational LLMs could enrich the PROs space, but incumbents with already sophisticated automation capabilities like PatientIQ may be better positioned to simply integrate LLMs into their already comprehensive offerings.
Conclusion
Patient panel management is a highly administrative, busy and yet important task. This space is being transformed by startups using technology in innovative ways to improve both provider and patient-initiated outreach, asynchronous care, pre-op and post-op care, both general and targeted (e.g. seniors) population health management solutions, and software solutions for PROs.
In summary, we see an opportunity for LLMs to make major impacts in several areas with some being more easily addressed by startups and others by incumbents:
Startup Opportunities:
LLM-enabled asynchronous low-acuity care: low-acuity conditions typically involve repetitive, guideline-based care. Possible solution: LLM to synthesize medical record, LLM to ask personalized medical questions, and medically fine-tuned LLM to suggest plan of care. Brief review by provider when necessary.
Why a Startup Opportunity? This may be too complex of a solution for a distracted incumbent to develop as an internal feature. Less encumbered and more focused startups may fare better.
Toss-Up Opportunity (Startup or Incumbent):
N/A
Incumbent Opportunities:
LLM-enabled system to handle patient- and provider-initiated outreach: essentially automating most of patient panel management. Combine traditional data analytics and AI with LLMs to proactively identify and chat with at-risk patients and share guideline-based recommendations reviewed by the appropriate level of provider when necessary.
Why an Incumbent Opportunity? Incumbent population health companies have a major advantage, since their existing customers, access to patient data, and data analytics give them a head start.
Outpatient Summary Figures
Thank you for joining us on our journey through outpatient care. Below we have included 2 summary figures:
Figure 1: A Guide to Who Wins: Startups or Incumbents
Figure 2: Generative AI Use Cases: Who Wins and How Quickly
Please see the captions below each figure for additional information.
Figure 1.
Figure 2.
 | A guest post by
|
 | A guest post by
|
Great article. A fan of journey mapping, I think the maps presented represent how things generally work today. Very helpful for internal discussions. Appreciate the references to the companies focusing on elements of the journey. The granularity is helpful. Look forward to the next two articles.