Future AI Opportunities for Improving Care Delivery, Cost, and Efficacy

With its ability to identify effective and economical services and treatments, artificial intelligence has the potential to create savings by eliminating waste in the system.

In the United States, healthcare spending represents one-fifth of the economy, with between one-quarter and one-half of all medical expenditures tied to waste.

Treatment variability, cost modeling, and drug discovery are three areas contributing waste to the system, but they are also opportunities for healthcare AI applications to help improve both quality and cost.

In 2017, Accenture analysts demonstrated the potential of artificial intelligence in healthcare to realize $150 billion in annual savings by 2026, with a focus on clinical health IT applications. Additionally, analysts pinpointed ten areas where AI applications were most likely to yield the largest savings:

• Robot-assisted surgery: $40B
• Virtual nursing assistants: $20B
• Administrative workflow assistance: $18B
• Fraud detection: $17B
• Dosage error reduction: $16B
• Connected machines: $14B
• Clinical trial participant identifier: $13B
• Preliminary diagnosis: $5B
• Automated image diagnosis: $3B
• Cybersecurity: $2B

The perceived benefits of healthcare AI are myriad, yet much depends on the support of providers, payers, and policymakers to ensure that any of these applications come to market and are adopted widely.

Applying forms of artificial intelligence to treatment variability, cost modeling, and drug discovery should lead to widespread changes in healthcare spending. Below we detail how AI will assist in turning these opportunities into cost savings and improved outcomes.

Standardizing care and reducing variability

A growing body of research shows that variations in treatment depend on numerous factors (e.g., disease, coverage, geography). And economists continue to cite treatment variability as a common form of wasteful healthcare spending.

“Waste in medical care comes in many forms. One clear cause is misallocated treatments: spending on care that is not clinically valuable or not spending on preventive services,” David Cutler, Otto Eckstein Professor of Applied Economics in the Department of Economics at Harvard University, wrote in Health Affairs.

Treatment variability is widespread and difficult to calculate. A recent study of expenditures and healthcare utilization for adults with low back and lower extremity pain found that “deviation from management guidelines is common and costly.” Patients treated according to guideline recommendations accrued lower costs.

A startling 1.2 percent of patients in the cohort of 1.3 million accounted for roughly 30 percent of total expenditures. Considering that 80 percent of Americans will experience low back pain or lower extremity pain at least once in their lifetimes, treatment variability will continue to lead to significant waste if left unchecked.

Early in 2019, the Food and Drug Administration noted the potential for artificial intelligence to reduce treatment variability and support the next generation of clinical trials. The federal agency pointed to present-day efforts by providers and researchers that tap into EHR and other data sources to train machine learning algorithms to identify optimal care pathways.

At the 2018 American Society of Echocardiography Annual Scientific Sessions, developers of a deep learning tool for cardiovascular imaging demonstrated just that. Their AI tool showed less variability in evaluating left ventricular ejection fraction (EF) than the average variability of cardiologists.

“By supporting fast, efficient and accurate AI-assisted echocardiogram analysis, the algorithms can allow physicians to focus on putting results into context for the patient — guiding prognosis and course of management,” said Richard Bae, MD, FACC, Director of the Echocardiography Laboratory at the Minneapolis Heart Institute and co-author of the study.

When applied to other diseases with high levels of treatment variability, AI tools could generate outsize savings by standardizing clinical decision-making.

Cost predictions and modeling

Artificial intelligence, particularly machine learning, is expected to play an important role in cost modeling by highlighting opportunities for reducing costs associated with specific diseases and treatments. With providers and payers engaged in more and more risk-based agreements, they have a growing need for timely and accurate data that can meaningfully inform their conversations about the total cost of care.

In 2018, the Society of Actuaries demonstrated the ability of machine learning to predict which patients were most likely to incur high costs by analyzing claims data on 47 million members between 2009 and 2015 maintained by the Healthcare Cost Institute.

Researchers found that just 17 percent of members included in the database were responsible for nearly 75 percent of all healthcare expenditures. Member cost history had the most significant impact on the probability of high costs. Prescription drug coverage and gender were also indicators of future high-cost patients, but researchers found that these characteristics have less significant effects on patient costs than member cost history. By addressing their most costly patients, providers could have an immediate impact on the total cost of care for these individuals.

Artificial intelligence, however, can be brought to bear on more specific care episodes and their associated costs. A graduate research project at the University of Pennsylvania showed the ability of a machine-learning algorithm to predict the cost of lung cancer treatment better than traditional cost prediction models, by 15 to 33 percentage points.

According to the author Jiaqi Li, “modern big data and machine learning tools are highly relevant to health care cost estimation.” Li concluded by noting that that “obvious advantage of machine learning algorithms over traditional statistical methods is efficiency: we see better predictive powers and efficiently reduced models.”

What’s more, artificial intelligence and machine learning offer the ability to test models without a randomized controlled trial and the ethical and feasibility implications of testing cost models on patients.

If artificial intelligence can identify high-risk, high-cost patients based on claims data alone, then it should yield even more critical insights when that data is combined with other sources of high-quality information such as the clinical record.

Drug discovery and development

Bringing a new drug to market is an expensive undertaking, with estimates hovering around $2.6 billion.

“A lot of that effectively goes down the drain, because it includes money spent on the nine out of ten candidate therapies that fail somewhere between phase I trials and regulatory approval,” according to a report in Nature.  

Not only are pharmaceutical companies in agreement that things need to change, but many are also engaged in AI-powered efforts to improve drug discovery. Pfizer, Sanofi, and Roche are three biopharmas active in using machine learning to develop effective drugs and therapies more efficiently.

“We are turning the drug-discovery paradigm upside down by using patient-driven biology and data to derive more-predictive hypotheses, rather than the traditional trial-and-error approach,” the head of biotechnology company Berg, Niven Narain, told Nature.

The challenge ahead for drug discovery is educating the researchers who will train these intelligent systems to identify unknown relationships at the genetic level. Data suggests that a sizeable number of drug-discovery researchers are unaware of AI uses in their field. In 2018, a global survey of 330 scientists working in drug discovery found that 41 percent were unfamiliar with AI uses in their field.

That lack of knowledge about artificial intelligence tools is detrimental to developing effective drugs and therapies. Last month, Bianca Nogrady of The Scientist reported that a graduate student at Washington University in St. Louis discovered the mechanism responsible for adverse responses to oral antifungal medications that led to fatal cases of liver failure in Australia.

Using a machine learning algorithm, Na Le Dang made an extraordinary discovery.

“Trained on large numbers of known metabolic pathways, the algorithm had learned the most likely outcomes when different types of molecules were broken down in the organ. With that information, it was able to identify what no human could: that the metabolism of terbinafine to TBF-A was a two-step process,” wrote Nogrady.

The discovery of two-step metabolites evaded traditional research but not Dang’s machine learning algorithm. It solved a longstanding “mystery” with lethal implications for users of terbinafine and highlighted the ability of AI to recognize previously unobserved patterns.

By applying artificial intelligence to drug discovery, the technology cannot only help in the creation of effective drugs but also increase the effectiveness of existing ones after the fact.

Artificial intelligence in healthcare is still limited, but the medical and financial gains to be made are too much to ignore. Forward-thinking healthcare organizations will come to recognize high-cost areas as opportunities to apply novel technology and maintain their competitive advantage over their peers.