*By Anthony Cardillo, MD*

  


Walking into a doctor's office or hospital in the United States, one can still find humming fax machines and flickering CRT computer terminals. A doctor can be urgently "paged" on a beeper or summoned by an overhead analog intercom. But despite the 1980s-era technology still on display, medicine is in fact experiencing an unprecedented artificial intelligence boom. 

# A Very Short History of Digitized Health Records

The Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 prompted [rapid digitization](https://journalofethics.ama-assn.org/article/hitech-act-overview/2011-03) within US health care. Financial incentives were offered in exchange for adopting Electronic Health Records (EHR), leading to the creation of digital datasets from a once paper-dominated industry.

Epic is one of several large corporations that began developing EHRs to replace paper records. Today, over 300 million people — nearly 4% of the global population — have a medical record originally created by Epic.

The US is heavily invested in electronic health care, with the Centers for Disease Control publishing regular reports on adoption by office-based physicians. The latest published report indicates that in 2019, 90% of office-based physicians used EHR — a decrease of 2% since 2018. Whereas the number of offices using a *certified* EHR, which has been tested and approved by the US Centers for Medicare and Medicaid Services, is lower, at 70%.

Over the last 10 years, I have worked in seven healthcare locations. Six of the seven used a certified EHR, and one relied on paper but was actively seeking an electronic system. I predict that the adoption rate will continue to rise, reaching approximately 90% by 2030. Almost every healthcare office I know is adopting electronic systems because they provide quicker results, aid with patient scheduling, and streamline the referral process.

# Investing in AI

With the structured data EHR make available, large tech firms began to invest significantly at the intersection of health care and AI. In 2013, IBM pivoted the Jeopardy-winning "[Watson](https://www.forbes.com/sites/bruceupbin/2013/02/08/ibms-watson-gets-its-first-piece-of-business-in-healthcare/?sh=2788caea5402)" AI to a new role as an oracle of medical knowledge at Memorial Sloan Kettering hospital where it was tasked with managing medical resources while providing lung cancer treatment recommendations. 

Epic created a machine learning model capable of notifying doctors when a patient is at risk of developing sepsis, a life-threatening condition causing organ failure. This alert is seamlessly integrated within the patient's electronic chart.

In imaging, numerous companies have developed AI able to read x-rays, CT scans, and MRIs to alert radiologists of potential risks. Though each of these AI readings requires a human to verify the output. 

By 2023, the FDA had [authorized over 500 AI/ML-enabled medical devices](https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices), with more than half of these authorized since 2020. At the current rate, one might expect 1,000 devices authorized by 2025. However, the FDA's regulatory posture toward "[Software as a Medical Device](https://www.fda.gov/medical-devices/digital-health-center-excellence/software-medical-device-samd)" is the key factor, and I anticipate a slowdown of approvals relative to the 2020-2022 boom. I predict we don't reach 1000 until after 2025. 

While highlighting innovations, it's important to note some significant failures in integrating biotech and AI into healthcare. In 2021, external validation efforts revealed that Epic's deep learning sepsis model [fell short](https://www.fiercehealthcare.com/tech/epic-s-widely-used-sepsis-prediction-model-falls-short-among-michigan-medicine-patients) of its advertised performance. Moreover, during the pandemic, the sepsis algorithm's frequent alerts led to counterproductive [alert fatigue](https://www.fiercehealthcare.com/tech/epic-s-sepsis-algorithm-may-have-caused-alert-fatigue-43-alert-increase-during-pandemic). Researchers at the [University of Pittsburgh](https://www.news-medical.net/news/20211214/Medical-AI-and-radiologist-experts-may-be-vulnerable-to-adversarial-attacks.aspx) also probed computer-aided diagnosis models using adversarial attacks that modified the contents of mammogram images, resulting in incorrect diagnostic outputs for nearly 70% of cases. 

# Aligning AI With Health Care's Triple Aim

Most stakeholders in the business and practice of medicine adhere to the "Triple Aim":

• improve the experience of health care

• improve the health of populations

• reduce the societal and individual costs of health care

The Triple Aim drives significant trends in hospital care, such as the "hotelization" of hospital rooms to improve experience, implementations of universal health care to improve population health, and the elimination of "fee-for-service" payment models to cut costs.

Because the Triple Aim necessitates tradeoffs, the future risks and benefits of AI largely depend on how it is employed to address these competing priorities. Enhancing patient experience with modern rooms, beds, and entertainment comes at a monetary cost. Rapidly optimizing population health metrics through focused public health initiatives can deprioritize individual health needs.

Taken together, competing stakeholder interests, lack of consensus on ethical principles, and insufficient efforts to quantify potential impacts are cause for concern that the US health care system could incubate AI applications that fail to balance these crucial trade-offs.

# Tomorrow's AI-Assisted Procedures and Diagnostics

AI is already used in [some laboratory diagnostics](https://pubmed.ncbi.nlm.nih.gov/33904171/), and we are fast approaching the day when doctor visits routinely include AI-assisted diagnoses and procedures. Currently, AI interprets patient x-rays and CT scans, analyzes pap smears for cancerous cells, recommends medications for cancer patients, diagnoses cancer shortly after surgery, and reports unusual practices to insurance companies for review. 

Numerous clinical trials and research efforts aim to expand medical AI functionality, including efforts to perform autonomous [procedures](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7901147/) and to [triage](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7861270/) limited resources—to name just two ethically fraught areas. 

Autonomous surgical robots can perform at the level of surgeons in certain highly technical procedures, and computer vision can identify cancerous lesions in real-time. When will the first fully autonomous surgery or procedure be performed? 

Although surgical robots like the da Vinci automated platform exist, no fully automated surgeries have been performed on humans. However, the fundamental technologies are available. I predict we will see such surgeries in the early 2040s, beginning with lower-risk procedures like appendectomies, which have a relatively straightforward protocol. 

# Conclusion

Medical AI holds the potential to make healthcare more efficient and effective, as well as improve patient outcomes. However, it is still in its early stages of development. There are many challenges to be addressed, such as eliminating bias, developing algorithms robust to edge cases, and carefully ensuring privacy and security throughout the entire AI pipeline. But if these challenges can be overcome, AI could revolutionize healthcare in ways once thought unimaginable.