Future of Healthcare: Personalized, Human-Centered & AI-augmented

In this second part of my series of blog posts on deep and frontier technologies (you can read my first part on protein engineering here), I touch upon the intersection of AI and healthcare, with a focus on painting an overarching picture from a European, early-stage venture lens. I will provide a high-level overview of the opportunity, the tailwinds, the problem landscape, the tasks ripe for AI disruption, the startups out there, challenges, and the emerging opportunities created by advances in generative AI and large language models (LLMs)—and what we are excited about. This is a long piece, so bare with me here—or just read the tl;dr below if you only have a moment.

TL;DR

If you only have a few minutes to spare, here is what you need to know about the intersection of AI and healthcare:

A generational shift in healthcare

• A $10 trillion problem. Healthcare is at a crossroads: while spending is soaring globally, health outcomes such as life expectancy are plateauing—or even declining, and the prevalence of disease is on the rise. Something needs to change.
• There is hope. Thanks to rapid advances in AI, particularly in LLMs, we are now on the brink of a generational shift in how healthcare operates. From automating back office workflows such as medical coding and billing to deciphering medical images at superhuman accuracy, AI is transforming all aspects of healthcare delivery.
• Ready for adoption. Coming out of the pandemic, healthcare providers that were once skeptical and resistant to previous waves of AI—as they typically overpromised and underdelivered—are now finally implementing solutions at a broader scale, recognizing that this time around AI can have a real impact on patient outcomes and the bottom line.

The future is personalized, human-centered & AI-augmented

• Personalized and full-context. We will no longer have to turn to the dark corners of the internet for self-diagnosis; instead, we will all have access to a personal AI health companion equipped with the full context of our health, from medical history to core body signals. This is the future our portfolio company Livv is building towards.
• Human-centered care. The days of manual note-taking into clunky systems are a relic of the past: startups like Corti and Nabla are eliminating doctors’ time spent on documentation by providing instant clinical notes, updating medical records, and analyzing the content—and ultimately giving back “the gift of time” to clinicians and allowing them to focus on more human aspects of care.
• AI-augmented. With OpenEvidence democratizing access to the vast universe of medical literature through a single query and startups such as Glass Health supporting clinical decision-making based on patient data, we are poised for a future where every doctor is equipped to make the best decision, grounded in evidence and augmented by real-time AI.

Coping with challenges

• Model accuracy. Deploying LLMs in healthcare is limited by their tendency to “hallucinate”, generate plausible but factually inaccurate information, and not preserve critical details. Startups like Xyla are working on aligning and grounding AI models in truth and making sure they are trained on high-quality, bias-free data.
• Security & trust. AI providers must address concerns around data security, trust, and patient confidentiality, while also accounting for ethical considerations, such as keeping humans in the loop and ensuring accountability.

While still in its infancy, AI is likely to fundamentally disrupt how healthcare operates—and in the process, enable healthier and longer lifespans for all of humanity. We as a fund are extremely excited about teams building at the intersection of AI and health.

Healthcare & AI—The Opportunity of a Generation

In a post-pandemic world, healthcare systems are under intense pressure due to inflationary pressure, shortage of staff, and the economic downturn. As healthcare spending continues to outpace inflation, budget cuts have shocked the health sector, resulting in a devastating impact on quality of care. We’ve all experienced it: we wait longer, it’s harder to access the right services, and there’s an overall decline in the quality of care. This is the new normal in Western healthcare systems.

However, there’s hope. Thanks to technological advances, particularly in AI, we are poised for a generational shift in how healthcare operates. Particularly, the rapid progress in Large Language Models (LLMs) is unlocking transformative opportunities across entire care workflows, representing an extraordinary leap forward in optimizing healthcare delivery. Ultimately, AI will enable us to give back the gift of time to doctors and allow them to focus on what matters: the real human interaction with patients.

It’s now possible to imagine a future where…

Everyone has their longitudinal health companion. Using data from wearables such as Oura, Veri, and Apple Watch, an AI health companion monitors our core body signals, detects any slight changes in our mood or health, and ties it into the full context of health: exercise patterns, medical records, lab results, and even genomic data. Through a unified interface, we can access insights, ask questions in natural language, and check symptoms. Consent provided, the system connects with our care provider and alerts  us if a severe medical need arises. Ultimately, such a companion can help us cope with social isolation, simulate our brains, or draft the perfect diet or exercise plan for us—while helping to facilitate the move to a more preventative health care system.

Care providers are augmented by an AI co-pilot. Gone are the days when doctors had to take manual notes of patient consultations, summarize them into clunky systems, or try to manually make sense of a patient’s medical history. Instead, a care provider has the full context of one’s health, can tap into the latest medical literature with a single query, and is augmented by an AI diagnosis and decision co-pilot, enabling more accurate care decisions for patients. Ultimately, clinicians can interact with all healthcare systems using natural language or voice dictation and request information or instruct an AI agent to perform a task.

Unlike in the past, AI's impact in these areas is not just about incremental improvements; it's about fundamentally reshaping how healthcare operates, making it more efficient, accessible, and human-centered. By leveraging the power of AI, the healthcare industry can address longstanding challenges and inefficiencies, opening the door to a new era of medical care that is more responsive to patient needs and capable of delivering high-quality care on a broader scale.

Therefore, I am extraordinarily optimistic that the application of computation and AI in healthcare is one of the greatest opportunities of the decade to create a meaningful and lasting impact.

Tailwinds

Significant VC funding interest 

Since 2010, venture capital firms have poured more than $44B into healthcare-focused AI and ML startups, according to data from Crunchbase. Although funding peaked at over $12B in 2021, spending in 2023 was still over $4B and above 2019 levels, highlighting the continued investor interest in the sector.

Advancements in AI & explosion of data

The advancements in AI, particularly in Large Language Models (LLMs), are unlocking unforeseen opportunities across the healthcare system, solving tasks from patient triaging to medical coding. Thanks to the sophisticated data processing and advanced natural language processing capabilities of LLMs, we can now interpret and generate human-like text, and extract insights from the exploding amount of healthcare data, enabling us to tap into the information stored in unstructured data, such as clinical notes and academic literature. The latest LLMs can, for example, already pass the medical licensing exam to become a doctor, as showcased by e.g. Google’s Med-PaLM model.

Readiness to adopt

Healthcare providers, once skeptical and resistant to the previous waves of AI, are now more receptive and eager to adopt new technologies. As necessitated by the pandemic, healthcare systems have adapted to virtual consultations, remote patient monitoring, and predictive modeling for capacity management. Alongside these shifts, healthcare providers have in recent years witnessed how hardware innovations are driving better patient outcomes through e.g. Intuitive Surgical’s surgery robotics systems—now expecting the same from their clunky software interfaces. More recently, we have all witnessed the capabilities of LLMs in our personal lives through e.g. ChatGPT, and are expecting similar capabilities from our work tools. Unlike in the past, where AI systems often overpromised and underdelivered, there's a growing recognition that the current landscape is markedly different, with AI now offering viable, impactful tools that can directly influence the bottom line.

Regulatory tailwinds

Over the past few years we have seen the European regulators pushing for important changes in EU regulation, paving the way for AI adoption. For example, back in 2019, the EU Commission presented a set of recommendations for the creation of a secure system that would enable EU citizens to access their electronic health records across member states. The starting point is making electronic health records accessible and exchangeable across borders, and standardizing the technical specifications to facilitate this. The shift to more accessible health data across member states is paving the way for startups to build the enabling layer for applying AI on top of the health record data. You can read more about the development of interoperable EHR systems across Europe here.

From paper-based systems to AI-augmented solutions

The ongoing disruption in healthcare mirrors how emerging markets leapfrogged from cash to mobile payments, bypassing traditional banking systems. Similarly, the healthcare sector is moving directly from outdated, paper-based systems to AI-augmented solutions, leapfrogging over traditional software adoption, representing a significant paradigm shift, aligned with the urgent need for more efficient, cost-effective, and patient-centered healthcare. The convergence of these factors—the pressures of a post-pandemic landscape, the rapid development of AI, and a newfound openness within the healthcare community—creates an ideal environment for the rapid adoption and integration of AI in healthcare.

Solving a $10 Trillion Problem

Rising costs as a percentage of GDP

Healthcare spending, as a critical component of GDP, has seen a dramatic rise both in the United States and the European Union, reflecting its growing impact on economies worldwide. According to The World Bank, healthcare spending in the US jumped from 12.5% in 2000 to a striking 18.4% in 2021, or a staggering $4.3 trillion. Similarly, in the European Union, government expenditure on healthcare reached a substantial €1,179 billion in 2021, accounting for 8.1% of GDP, representing a 33% increase from 2000, when healthcare spending was 6.1% of GDP, as reported by Eurostat, underscoring a steady upward trend in healthcare spending relative to GDP over the years.

Risings costs do not equate to better health outcomes

Rising healthcare costs as a share of GDP is not a bad thing in and of itself; however, if it does not equate to longer, healthier lifespans something is wrong. Despite the shock effect of the pandemic, we can see that the average expected lifespan in the US has plateaued since 2012. In the EU, the situation is better: healthcare costs before the pandemic stayed relatively stable—and even declined slightly-–between 2008 and 2018 while life expectancy rose substantially.  However, even in the EU, we have not been able to combat the growing prevalence of neurodegenerative diseases such as Alzheimer’s, cancers, or cardiovascular diseases, as indicated by the population-adjusted mortality from these causes of death; for example, global mortality from neurodegenerative has increased by over 60% in two decades.

Administrative burden and wasteful spending

The scope of the healthcare system has grown substantially in the past decades, leading to increased complexity and administrative burden. In the US, administrative costs are estimated to account for 15-30% of total expenditure, at least half of which does not contribute to improved clinical outcomes. Administrative complexities, duplicative services, and unnecessary treatments are major contributors to this wastage, accounting for approximately 25% of these expenses. Additionally, high drug prices and hospital readmissions further exacerbate the financial burden on the healthcare system. Such inefficiencies not only inflate overall healthcare costs but also divert resources away from initiatives that could enhance patient outcomes.

Increasingly fragmented and siloed care

There was a time when your family GP knew everything about your and your family’s health and access to care was more holistic and unified. Now, with the increased complexity of the healthcare system, the delivery of coordinated and efficient care is challenged by a significant level of fragmentation. This lack of synchronization is leading to preventable mistakes, from diagnosis to treatment execution. For example, the siloed nature of healthcare data—often confined within specific healthcare entities or systems—hampers the holistic understanding of patient health across stakeholders, often leading to the duplication of services, increased costs, and potential gaps in care.

Shortage of staff; burned out and depressed health workers

The healthcare sector is currently facing a critical staffing crisis with the EU alone facing an estimated shortage of 1 million health workers, challenging the provision of high-quality care for the continent’s citizens. The dire situation is further exacerbated by the increasing prevalence of depression and anxiety among healthcare workers: around 40% are battling depression and anxiety, with a staggering 70% experiencing symptoms of burnout. Contributing factors to this alarming trend include low retention rates, deteriorating working conditions, and significant unrest among healthcare professionals, evidenced by widespread strikes, including the largest NHS strike action on record in the UK, and similar actions in Finland, for example.

Big Tech & Healthcare

Big Tech has attempted to take a foothold in healthcare for almost two decades

Tech incumbents have had long-enduring ambitions to enter the healthcare market—but with mixed results. America’s five big tech companies—Alphabet, Amazon, Apple, Meta, and Microsoft—have poured billions into various healthcare investments. From Amazon’s online pharmacy to a failed Amazon-JPMOrgan-Berkshire joint venture to provide better employee health outcomes.

Apple has been tracking people’s health since the launch of Apple Health back in 2014, to provide its users meaningful insights into their health. Apple also launched health records back in 2018, allowing users to view, manage, and share their medical records—Google attempted the same already back in 2008, wound it down in 2012, resurfaced the service in 2018 as Google Health, and again dismantled the service in 2021. Similarly, Microsoft had its proprietary health record system, HealthVault, launched in 2007, but closed down in 2019. These early attempts by Big Tech to take ownership of patient data forced people to consider whether they could entrust Big Tech to take hold of their data; for many, the answer was no.

“If you zoom out into the future…and you ask the question, ‘What was Apple’s greatest contribution to mankind?’ It will be about health.” - Tim Cook, CEO of Apple (2019)

Advances in AI are sparking a renewed interest

The recent advances in LLMs have sparked a new emergence of initiatives by big tech to enter the healthcare space. Earlier this year, Google announced that its cloud business is working with Mayo Clinic to provide test access to its Enterprise Search on Gen AI Builder. In the context of healthcare, this LLM-powered chatbot builder enables care providers to draw insights from health records, imaging data, or lab results with a simple query. In May, Google also announced Med-PaLM 2, a medical LLM purpose-built to provide accurate and safe answers to medical questions. In 2022, Google also announced its Medical Imaging Suite to accelerate the development of AI for medical imaging by making data accessible, interoperable, and useful.

Meanwhile, Epic Systems, a leading medical record software provider in the US, is expanding its strategic collaboration with Microsoft Azure’s OpenAI service “to increase productivity, enhance patient care and improve the financial integrity of healthcare systems globally”. Similar to Google’s Enterprise Search, the collaboration will bring natural language query access to Epic’s vast trove of health record data, enabling care providers to explore data more intuitively. Additionally, Oracle Health (Oracle’s healthcare platform) is incorporating LLMs into their current product deployments and bundling them into their existing products pushed through their well-established customer bases.

AI Waves in Healthcare

We have now entered the fourth wave of AI in healthcare, an era defined by the progress in large language models and advances in natural language processing. While in the first two waves, innovation mainly came from large industry incumbents, such as Siemens, GE, Philips, and even IBM, two of the most recent AI waves are markedly different: we have seen an explosion of VC-funded healthcare AI startups founded since 2015. Startups founded after 2015, such as PathAI, Paige, Abridge, Unlearn, and Ambience have all raised more than $100m. With the democratization of foundation models and a significantly lower barrier to building AI products for healthcare, I believe that the vintage of companies from this era will disrupt the industry for decades to come.

The diffusion of AI in healthcare

As highlighted in previous sections, many healthcare tasks are well suited for disruption by AI solutions. To better understand the landscape of these tasks, I have mapped them out across two dimensions. Firstly, the different tasks are spread across a spectrum of patient-facing (”front office”) tasks vs administrative (”back office") tasks. As the name implies, the patient-facing vs. admin axis illustrates how close to the patient the task is. Secondly, the tasks are mapped along clinical vs. non-clinical, differentiating between tasks currently borne by specialized professionals, i.e. doctors, and tasks that can be completed by a non-medical professional.

I would argue that the diffusion of the current wave of AI across the back vs. front office dimension starts with the patient-facing applications; it is fundamentally easier to develop and implement an AI chatbot that can help patients check their symptoms (e.g. Ada) vs. building out a back office automation workflow that needs to connect and interact with a multitude of legacy systems where often staff needs to be trained.

Company mapping

Patient consultations & doctor co-pilots

For every hour a doctor interacts face-to-face with a patient, nearly two additional hours are spent on administrative work, including EHR and desk work. Startups on both sides of the Atlantic are scrambling to build the best care-provider co-pilot out there to reduce clinicians’ administrative burden, typically starting with scribing doctor-patient interactions in real-time. AI can then, for example, synthesize these conversations, augment them with EHR context, coordinate the next steps on the patient journey, make sure appropriate medical coding is completed, and provide a summary for the patient and their family—with the ultimate goal of giving back doctors the “gift of time” and making patient consultations more human-centered.

Our portfolio company Corti (recently raised a $70M Series B led by Atomico & Prosus), based in Copenhagen, is one of these. With a unique dataset and AI trained on 16m+ patients and 550k+ hours of unique consultation audio, Corti is an AI system that augments, automates, and analyzes patient consultations, and combines it with AI-based medical coding and triaging solutions. Another European startup in this space is Nabla, based in Paris.

Clinical search engines

AI-powered clinical search transforms the traditional methods of navigating through medical literature and research. It enables healthcare providers to input natural language queries and receive relevant, evidence-based answers in real-time from identifying the latest treatment guidelines for a specific condition to sifting through patient outcome data. A frontrunner in this space is OpenEvidence: ask a question such as “How to diagnose and treat West Nile virus”, and it will summarize, analyze, and synthesize the universe of peer-reviewed medical studies to provide you with an answer, grounded in truth.

Foundation models

General-purpose language models are typically pre-trained on a common crawl of the internet, some of which is incorrect or misleading, posing challenges for accuracy-critical domains such as healthcare. For example, Hippocratic AI has invested heavily in acquiring evidence-based healthcare content and trains its model using reinforcement learning with a human feedback process to validate the model with healthcare professionals. Using a slightly different approach, OpenEvidence’s AI model, powered by Xyla, is trained on high-quality medical literature and can provide answers grounded in peer-reviewed sources. Last year, Google also announced Med-PaLM 2, a medical LLM purpose-built to provide accurate and safe answers to medical questions.

Decision & diagnosis support

Taking a step further, companies are providing context-aware decision support systems. By integrating patient data with the latest medical knowledge, these AI-driven systems can suggest personalized treatment options, predict patient outcomes, and even identify potential risks before they become problematic. For instance, an AI model could analyze a patient's electronic health records (EHR), alongside current research, to recommend a treatment plan that maximizes efficacy while minimizing side effects. For example, Glass Health offers an AI platform for developing differential diagnoses and drafting clinical plans based on a patient summary.

Clinical datasets

To train LLMs tailored to the medical domain, accurate, high-quality training data is required. To aggregate such data across health systems, companies like Dandelion Health partner with health systems to provide de-identified, high-quality clinical data with ground truth clinical outcomes and expert labels. Some startups are working on creating synthetic datasets. For example, Syntegra has developed what it calls its “Medical Mind”, an AI model that understands the statistical distribution of all types of healthcare data, and then uses this information to generate synthetic patient records. Similarly, Unlearn has developed a digital twin-generating AI model, that ingests data collected from each patient and generates a digital twin, which can then be used to provide comprehensive, longitudinal predictions of potential outcomes.

Medical imaging

For a long time, AI has been helping to classify disease from radiology images, MRI scans, and electrical signals, thus helping to provide more accurate diagnostics, while saving costs and preventing misdiagnosis. Although AI for medical imaging has been around for a long time, it is easy to overlook how the current generation of AI can accelerate the speed of interpretation and precision of diagnosis. Startups are building solutions ranging from an operating system for radiology AI workflows (e.g. deepc) to improving and de-noising image quality (e.g. Subtle Medical) and improving diagnosis accuracy (e.g. Floy). Some European companies in this field include MVision, Incepto Medical, Gleamer, and Deemea.

In the broader field of medical imaging, current technology is enabling new forms of examinations that were not possible before. The noise-to-signal ratio in some applications used to be too poor to be able to generate anything meaningful. Now companies such as Neko Health are transforming medical check-ups for example. Another interesting domain is liquid biopsy, a medical test that detects cancer cells or pieces of DNA from tumors circulating in a patient's blood. AI contributes to various aspects of liquid biopsy as it allows for analyzing complex data from blood samples with high sensitivity and specificity. AI can also sift through the vast amounts of genetic data to identify relevant mutations for cancer diagnosis. However, as many of these companies do not necessarily have an AI-first approach, they have been excluded from the company mapping.

Patient monitoring

Patient monitoring is the continuous or periodic observation, recording, and analysis of a patient's vital signs or other health parameters over time to detect any deviations from normal health conditions early and make informed decisions regarding treatment and care. AI systems are capable of continuously monitoring vital signs and other health parameters in real-time, such as heart rate, blood pressure, oxygen levels, and temperature, analyzing this data more quickly and accurately than traditional methods, enabling early detection of potential health issues. For example, a Copenhagen-based startup, Teton.ai, is building technology for patient monitoring using computer vision to enable care staff to minimize hospital injuries, support staff’s ability to care, and help them make more data-driven decisions for operational efficiency.

Data structuring & processing

AI is well-suited to structure and process the vast amounts of unstructured and structured health data—ranging from patient records, clinical notes, and imaging data, to genetic information—into usable formats for better decision-making and insights. LLMs can extract relevant information, identify patterns, and interpret complex medical jargon from diverse data sources, transforming them into structured data that is searchable and analyzable—and thus ultimately helping health care professionals better diagnose, monitor, and prevent diseases. For example, Spain-based IOMED has developed an NLP system that unlocks data from structured and unstructured data sources.

Medical record retrieval & analysis

As medical records continue to be distributed across different siloed Electronic Health Record (EHR) systems. Accessing a patient’s comprehensive medical history, treatment, and care over time remains a challenge. In some countries around the world, the government has taken steps to ensure all medical records—from private and public care providers–are stored in one place, such as Finland’s Kanta system, but this is not the case for most countries. As mentioned earlier, the EU is pushing for better interoperability across EHR systems, but it remains to be seen how effective that will be. Meanwhile, startups such as meMR, based in New York, and are working to create a fully automated, full-service record retrieval platform, transforming and categorizing the data into one holistic data source.

Patient interactions & care coordination

As part of the patient journey, care providers interact with patients online. To automate these interactions, startups are building conversational AI healthcare assistants. This can include virtual triaging, where patients can interact with a conversational AI chatbot that helps them check symptoms, self-triage, receive care guidance, and schedule appointments. In addition to such services, Clearstep is also automating and streamlining the post-discharge care journey and interactions by, for example, making sure patient monitoring programs are personalized and optimized.

Call center operations

AI is well suited to streamline and manage call center functions in healthcare by automating the processes of handling patient inquiries, scheduling appointments, providing information on healthcare services, and offering support for billing and insurance queries. By leveraging AI, healthcare providers can offer round-the-clock assistance, enhancing patient experience and operational efficiency. For example, Birch AI classifies a patient call, writes a summary, and completes key fields in the system of record. Infinitus takes it a step further and fully automates complex calls in healthcare by making the calls automatically and collecting upwards of 150 data points per call “with superhuman accuracy”.

Consumer-facing AI

I have long imagined a future of healthcare where everyone has a longitudinal health companion. A companion that has access to your comprehensive medical record—maybe even your genomic data, and the real-time data from your wearables and other health tracking devices. It knows your baseline and can detect subtle changes in your physical and mental health. That is the future our portfolio company Livv is building towards, starting with medical records. A multitude of startups out there are building health companions and apps for consumers, from relationship AI chatbots (Meeno), and from personalized health tracking & coaches (Woop) to AI doctors accessible via a chat interface (Docus, Ada). In the market mapping, I have only considered a few example startups that have a direct link to the healthcare system.

Medical coding, billing & insurance

The typical patient journey from a medical consultation to billing is complex and varies by healthcare system. Typically, the process starts with the doctor making a diagnosis and treatment plan based on the patient consultation, including prescribing potential medication, ordering further tests, or recommending follow-up visits, all of which are stored in the patient’s medical record. The medical coders translate the services provided during the consultation, including diagnoses and treatments, into standardized codes. These codes are then used to communicate medical information between healthcare providers and payers (typically insurers or the government). Then the billing “department” creates a bill (or a claim) for the services provided, which is then sent to the patient’s insurance company or directly to the patient if the patient is paying out-of-pocket. The insurance company processes the bill and bills the patient on any portion of the services not covered under the plan.

Given the intricate nature of these procedures and the need for precision and compliance with various regulations, these tasks can be time-consuming and prone to human error. Integrating AI, especially Large Language Models (LLMs), into these processes offers significant benefits by automating tasks, enhancing accuracy, and improving efficiency. That’s why startups like Corti, Synaptec, and Slingshot, among others, offer products to automate medical coding with AI. Startups are automating various aspects of the billing and insurance workflow: Alaffia is working to automate pre-payment claim reviews, Pledge is automating insurance benefit verifications and patient out-of-pocket cost estimation, and Cair is automating the entire insurance workflow.

Operations & staffing

Within the back office of healthcare there are many other problems ripe for disruption by current AI technologies that are not captured in the above categories. For example, AI can assist in staffing by analyzing trends in patient flow and predicting peak periods, enabling optimal staff scheduling. It can also identify skill gaps and recommend training or hiring strategies to ensure the workforce is well-equipped to meet patient needs. One of the companies solving this is Anima which, among other things, offers staffing management and booking management. On the operations side, Latent Health is building generative AI for healthcare operations starting with pharmacy operations. By enhancing back-office operations and staffing efficiency, AI can contribute to more responsive, cost-effective, and high-quality healthcare services.

Challenges

Regulatory compliance

Startups offering services in the old continent have to be mindful of three key EU regulatory frameworks (among others): the medical device regulation, the GPDR regulation, and the EU’s AI act. The medical device regulation governs products with direct or indirect medical purpose that guide patient care decisions. GPDR is focused on privacy and data security, for example, mandating that data processing should only happen with explicit consent and for specified purposes. Additionally, GDPR empowers individuals with rights over their data, including access, correction, and deletion rights. The EU AI Act classifies AI systems according to the risk they pose to users, and establishes obligations depending on the associated risk level. For example, high-risk AI systems (medical devices) need to be assessed before being put into the market and also throughout the lifecycle.

Interestingly, we are seeing some startups building a tech approach to compliance. For example, Delve is building automated infrastructure to allow for HIPAA-compliance in days, and Scarlet is building a framework that provides manufacturers and regulators with a way to build and surveil medical device products in a continuous and scalable way.

You can read more about the EU’s AI act here, and the World Health Organization’s perspective on regulatory considerations on artificial intelligence for health here.

Hallucination and preserving the integrity of data

A challenge that arises in real-world clinical settings is the phenomenon of 'hallucination', where AI models may generate plausible but factually incorrect information. Large language models (LLMs), despite their sophistication, are lossy by nature. They excel at producing coherent text, but this doesn't always equate to accurate text, leading to a potential loss of nuanced details essential in clinical practice. Therefore, preserving data integrity, and making sure output is grounded in truth and linked to the source, are critical to mitigating hallucinations and ensuring the reliability of their outputs in healthcare applications.

Quality of data and avoiding of bias

Ensuring the quality of data feeding into AI models is pivotal to their performance, adhering to the principle of "garbage in, garbage out." Access to high-quality, relevant data is therefore a cornerstone in building a solid foundation for any AI system. To combat this, our portfolio company Corti has cultivated an exclusive dataset from 550k+ hours of unique consultation audio, an invaluable set of data for training robust AI solutions. Furthermore, as these models depend heavily on the data they are trained on, they are susceptible to biases if the data lacks representativeness—a particularly acute issue in clinical and medical domains. Therefore, providers of AI systems must ensure that models are trained on data free from gender, social, racial, or religious biases to avoid perpetuating inequalities.

Keeping data up to date (interoperability & data access)

Ensuring that data remains up-to-date presents a considerable challenge in healthcare given the necessity of managing data flows across the fragmented and siloed set of providers and systems. The initial acquisition of, for example, EHR data is a complex task in itself, and keeping the data constantly updated adds another layer of complexity. Startups need to integrate with the myriad of existing systems, each with its data standards and protocols. In response to these challenges, solutions like Co:Helm's "interoperability engine" have emerged, designed to facilitate the smooth exchange of healthcare data and bridge the gap between different healthcare IT systems, ensuring that patient data is not only accessible but also consistently updated across the healthcare ecosystem.

Data security, trust, and patient confidentiality

Deploying AI into the healthcare domain highlights critical considerations surrounding data security, trust, and patient confidentiality. Ensuring robust data security measures, strict adherence to enterprise security standards, and compliance with rigorous regulations such as HIPAA and SOC 2 are foundational in protecting sensitive patient information. It’s not about just safeguarding data; it's about fostering trust between patients and healthcare providers, assuring them that their personal health information remains confidential and secure in an increasingly digital world powered by AI.

Ethical considerations and human-in-the-loop

Moreover, the deployment of AI in healthcare—through decision support systems like Glass Health and OpenEvidence—necessitates a human-in-the-loop approach to navigate the complex ethical landscape of AI-augmented decision-making. This strategy ensures that AI's potential to improve healthcare outcomes is balanced with the need for reliability, accountability, and liability, aligning with human conventions, ethical standards, and legal requirements. Patients' concerns about the accuracy and impersonality of AI in their care underscore the importance of maintaining a human touch in interpreting and validating AI-generated insights.

The paradox of information

The advent of AI and human-centric healthcare systems, which grant patients unfettered access to extensive health information, risks overwhelming patients with information, potentially leading to increased anxiety and stress. Similar to how the ubiquity of social media has contributed to a rise in mental health issues by perpetuating comparisons and unrealistic expectations, the flood of health data can lead to hyperawareness and misinterpretation of one's health status, fostering unnecessary worry. Addressing this requires careful consideration of how information is presented to patients, ensuring it supports informed decision-making without contributing to the growing anxiety epidemic, especially among the younger generation.

Future Predictions & Our View

The future of healthcare is:

1. Human-centered care

Fueled by increasing health awareness and the democratization of health tracking and access to knowledge, the future of healthcare will be radically human-centered (or patient-centered but I prefer to avoid using the word patient). As people become more informed and involved in their health, they naturally demand more from their healthcare systems, advocating for care that prioritizes individual preferences, needs, and values in clinical decision-making.

The shift towards human-centered healthcare is supported by advancements in AI, which is already revolutionizing the way care providers interact with patients. By automating routine tasks of doctors and analyzing patient data more effectively, AI is making the doctor-patient interaction more human-centered. Ultimately, AI will give back to healthcare professionals the “gift of time”, enabling them to focus on the more human aspects of care—such as empathy, understanding, and personalized interaction.

We are poised for a future where healthcare is not just about treating symptoms but holistically supporting individuals in a way that is most meaningful to them, ensuring that technology serves to enhance, rather than replace, the human-centeredness of medical care.

2. Personalized and full-context

The future of healthcare is gearing towards radical personalization, transforming the patient experience into one that is intimately tailored to each individual’s health journey and based on the full context of one’s health. We will all have access to a personal health companion, and will no longer have to turn to the dark corners of the internet for self-diagnosis and help.

Such a companion will have all relevant context on our health, spanning from our comprehensive medical history and genomic data to the previous night’s HRV and body temperature (e.g. Oura), and from exercise trends (e.g. Strava) to your glucose levels (e.g. Veri). The companion will detect and alert on any abnormalities, answer any questions we might have about our health, and help us self-diagnose symptoms—and seamlessly coordinate a consultation when needed.

If given consent, doctors will have access to the same data in a separate interface, allowing for unprecedented precision in diagnosis and treatment. Powered by these advancements, the personalization of care will ensure healthcare is not only more responsive but also pre-emptively attuned to the individual condition of a patient, allowing for a leap towards truly preventive and personalized care.

3. AI-augmented decision-making

With tools like OpenEvidence, physicians will have immediate access to the vast universe of medical research, clinical guidelines and medical trials, continuously updated and rooted in evidence-based practices. Integrating longitudinal health data and outcomes for specific treatment scenarios will ensure that care decision are grounded in truth. 

Additionally, doctors will have access to a real-time AI co-pilot equipped with the full context of a patient’s health. A co-pilot that will not only ambiently take notes and summarize interactions but more importantly, they will ensure and nudge doctors to ask the right questions and help them make informed decisions in real-time, enabling precise decision-making for every patient. Ultimately, this will improve patient outcomes and drammatically reduce medical errors.

4. “Command Line”

Lastly, I believe the future of healthcare is poised for a radical disruption in how doctors interact with healthcare systems. As inspired by Morgan Cheatham’s great article on “Command Line Medicine”, we will transition from cumbersome graphic user interfaces to a more human-centered interaction with systems—to command line medicine. Eventually, healthcare providers will have access to a streamlined interface that is connected to every healthcare system, allowing care providers to access and interact with healthcare systems efficiently.

Through natural language processing, doctors will be able to request information (”read”) and instruct the system to perform a function or a task (“write”) using either voice dictation or a written query—or maybe even interact via a brain-computer interface in the future. Through such a system, doctors will have instant responses to questions like “What have the exercise patterns of the patient been over the past three weeks?”, and instruct the system to, for example, “Update the patient's progress note and write a prescription for Clopidogrel”.

This transformation, augmented by AI, will offer care providers a profound level of efficiency and precision, enabling them to focus on providing personalized, human-centered care, while signaling for a future where technology abstracts away system complexities, allowing instead for more human interaction with computational systems in healthcare.

Conclusion

We are on the lookout for startups building at the intersection of AI and healthcare, especially ones that are solving problems from the European/EU perspective. We are tremendously excited about the opportunities in the space and truly believe there is a paradigm shift happening in healthcare right now. So, whether you are just exploring ideas or have already found your product market fit, we would love to hear from you.

.

.

.

Selected Readings

Selected articles

Command Line Medicine by Morgan Cheatham (Bessemer Venture Partners)

Generative AI That Saves Lives by Kahini Shah (Obvious Ventures)

The Next WebMD: An LLM as Your Front Door to Healthcare Daisy Wolf and Julie Yoo (A16Z)

Where Will AI Have the Biggest Impact? Healthcare Daisy Wolf and Vijay Pande (A16Z)

Commercializing AI in Healthcare: The Jobs to be Done by Jay Rughani, Daisy Wolf, Vijay Pande, and Julie Yoo (A16Z)

Health care financing in times of high inflation by OECD

Generative AI for Healthcare Perspective by Josephine Chen (Sequoia)

The 3P's of AI in Healthcare by Nikita Andersson & Alexander Jenkins (Hummingbird VC)

Software-First Care & The Next Generation Of Healthcare Technology by Steph Weiner (FirstMark)

What is the future of healthcare by Julie Yoo (A16Z)

To Make a Real Difference in Health Care, AI Will Need to Learn Like We Do by Vijay Pande (A16Z)

These Doctors Aren’t Sweating AI — Yet by Alvin Powell (Harvard Medical School)

The Role Of Administrative Waste In Excess US Health Spending

Government expenditure on healthcare by Eurostat

Health and care workforce in Europe: time to act by the World Health Organization

Europe Is Struggling to Keep its Health Systems Afloat by Stefan Anderson (Health Policy Watch)

Selected academic publications

The promise of large language models in health care

Large language models propagate race-based medicine

Large language models in medicine

The Utility of ChatGPT as an Example of Large Language Models in Healthcare Education, Research and Practice: Systematic Review on the Future Perspectives and Potential Limitations

Attention is not all you need: the complicated case of ethically using large language models in healthcare and medicine

A large language model for electronic health records

Leveraging Generative AI and Large Language Models: A Comprehensive Roadmap for Healthcare Integration

‍

‍