Combining the Power of Artificial Intelligence with the Richness of Healthcare Claims Data: Opportunities and Challenges

A significant advantage of AI over standard statistical analysis is its ability to examine large multidimensional data with many variables. A statistician may easily be overwhelmed by a large number of potentially usable variables. AI algorithms, in contrast, are often able to identify which variables are important and detect the optimal combination of these variables for the task at hand.

By combining claims data with other information sources, such as patients’ laboratory results and EMRs, the knowledge and experience of a wide range of medical practitioners can be pooled together to detect complex patterns indicative of certain illnesses. The identification of these patterns can improve the early detection of diseases and the detection of underdiagnosed or rare diseases, provide more accurate diagnoses, or lead to a more proactive offering of personalized preventive services.

In the case of conditions that are underdiagnosed, for example, a sample of patient claims data (combined with laboratory values, specialist visits, and physician notes) can be analyzed by physicians to determine which of the patients should have been diagnosed and at what point in time. This process can then be used to train algorithms to detect other instances that should be flagged and identify early markers and indicators of future disease onset. For example, recent research has used AI to predict type 2 diabetes mellitus using claims data [11]. More generally, there have been a number of applications using AI with other types of data (e.g., CT scans, or genome sequencing in precision medicine) to support diagnoses, either through the prediction of disease onset or the identification of drug-resistant strains of disease (e.g., drug-resistant tuberculosis [12]).

Another promising application of AI is the detection of treatment adverse effects. Adverse effects are typically hard to discover in standard human trials because of relatively small sample sizes. Large claims data may help discover these adverse effects, in particular when combined with the pattern recognition power of AI. Recent research has used machine learning algorithms to detect adverse drug reactions using EMR data [13]. Healthcare administrative data can be linked with EMR data to add more clinical information [2], and may improve detection. More generally, as discussed by Dadson et al. [14], AI has been used to detect drug–drug interactions based on the text in government and scientific reporting systems and may be used with genomic data to determine genetic factors associated with particular adverse effects.

AI has also been used broadly to systematically learn from past experience. Examining patterns in pooled data can improve the monitoring of services, such as hospital readmissions or diseases contracted on site. For example, machine learning analysis on patients’ electronic health records, demographics, medical histories, admission and hospitalization histories, and likely exposure to the Clostridium difficile bacterium was shown to increase the accuracy in predicting the risk of hospital infections [15].

AI also has the potential to identify and reduce the effects of common biases in healthcare, such as doctor biases and omitted variable bias.

AI has had an influence on each step of the processing pipeline for insurance coverage: claim submission, claim adjudication, and fraud monitoring. Its influence is likely to continue well into the future.

An efficient and accurate intake process for claims submissions can reduce burdens on patients and the entire medical system. For example, AI image recognition can be used to facilitate the process of submitting and coding claims. One insurance company has trialed a procedure to automate claim approvals. Patients electronically submit photos of their hospital bills and, within moments, receive notification of receipt, approval, and credit to their account [25]. The cost savings associated with this automation process may allow insurance companies to scale up and provide more coverage. Some insurance companies have developed algorithms that recommend and incentivize healthy habits and behaviors to policyholders, such as exercise and nutrition strategies [26]. If successful, these may avoid claims submissions for preventable illnesses altogether.

Claim adjudication involves checking coverage, limits, contracts with providers, pharmacies, appropriate diagnosis, and procedure coding. Complex and suspicious claims trigger secondary, usually manual, processing steps. AI can save time via the efficient processing of normal claims (e.g., through increased automation in the settlement of claims based on their complexity and known patient history) [27]. In addition, AI can be used for early detection of abnormal price patterns among healthcare providers. This may help insurance companies monitor their costs and understand whether price increases reflect actual quality improvements.

Increasing the accuracy of triaging claims through AI can reduce losses due to fraud (e.g., a provider billing for services that were not actually provided or the provision of unnecessary medical services to generate insurance payments). While the share of actual fraudulent claims is generally very low [28], these can add up to very significant dollar amounts. Just counting savings to Medicare, the Health Care Fraud Unit of the US Department of Justice’s Criminal Division recovered and transferred billions of dollars in the 2017 fiscal year [29]. Several data mining efforts have been undertaken with increasing levels of sophistication to improve fraud detection [30]. A Google Patent search for “healthcare machine learning fraud detection” reveals over 7700 results [31]. The general challenge is to limit false positives, which can lead to losses in reputation of targeted firms as well as enforcement efforts that are channeled inappropriately, while increasing the proportion of fraud detected. The rules to detect individual fraud are now relatively simple to implement. Recently, more complex network fraud pattern recognition algorithms have been developed with the potential to significantly reduce the costs for insurers [32].

Most ethical problems related to confidentiality can be dealt with by anonymizing the data. In the past, removing social security numbers, phone numbers, names, and most date of birth information was sufficient to anonymize data. However, the recent emergence of third-party datasets (e.g., social media and cell phone tower access records) has provided additional avenues to triangulate datapoints and identify individuals within the data [33]. This creates additional challenges and a potential need for better de-identification techniques that ensure that shared information cannot be used to approximate personal information even when combined with pattern detection techniques.

As is the case with standard statistical methods, AI algorithms built on biased data will lead to biased predictions. Supervised machine learning algorithms, for instance, learn relationships between outcome and predictor variables after being trained on a set of examples for which the true outcome is known. Algorithms will have trouble classifying new examples that are very different to what they have seen before. In machine learning this problem is referred to as “hasty generalization” [34]. While this problem is not unique to AI, the rapid speed at which individuals are now able to analyze large datasets has raised concerns that some believe the need for evidence of causal relationships is unnecessary, large databases by themselves are enough, and that the numbers will speak for themselves [35]. While complex correlation patterns are useful for making predictions, they may be of limited use for explaining why certain events occur [36].⁷ The importance measures [8, 9] and highly interpretable AI models [10] discussed earlier can be useful in this regard.

Using AI or any statistical tool requires careful consideration of both the underlying data and the results. There are inherent biases in most claims databases due to sample selection (e.g., the characteristics of the insured population and varying standards of care), such that results may not be generalizable to other population groups. For instance, evaluating the effect of a drug that requires a very demanding treatment on a subpopulation with a high adherence rate may overestimate the effectiveness of the drug on the general population. In addition, the decisions made in real-life applications based on an algorithm’s results can have lasting effects (e.g., if an individual is incorrectly declined insurance because their profile shares some characteristics with at-risk groups) [39].

Supplementing claims data with information from other sources may limit the potential bias from unobserved variables. As discussed earlier, particularly rich data may also be used to triangulate and proxy for information that is still missing. However, rather than relying on technology to solve potential issues ex post, recent research has investigated whether technology can help gather more exhaustive, rich, and less biased datasets ex ante. For example, the blockchain technology associated with bitcoin is being considered as a potential solution to securely store healthcare data. Individuals would have ownership over their own personal data and would potentially receive financial incentives to make it widely available. Blockchain technology could also help document the provenance, veracity, and selection of the data used in studies. This could greatly reduce the widespread difficulty in replicating published medical research. Relatedly, this documentation could reduce the ability to cherry-pick results [40].

Just as AI algorithms can help detect a disease in its earliest stage, they can also help predict future health costs and therefore be used, in theory, to adjust insurance premiums based on combinations of personal characteristics. If an insurer can predict illness reliably, competition could lead to lower premiums for predictably healthier individuals and higher premiums for predictably sicker ones, even if the characteristics used for prediction are beyond an individual’s control.

Even if such a process could be appropriately implemented, it may be economically inefficient. An individual who is uncertain about their future sickness would want to smooth their future income across possible states of nature (e.g., sick vs. not sick). The individual would want insurance, but if the insurance company has reliable forecasting technology, competition could undermine that person’s ability to obtain insurance, leaving the individual worse off.⁸

The extent to which such insurer strategies are already used and how much should be prohibited by law is often the subject of debate, particularly since similar issues appear in other industries. For example, research on credit access has examined the use of variables correlated with race and gender when evaluating loan requests and the potential bias against certain groups [42, 43].

At issue in these cases is not the methodology itself but rather its use. Let us assume, for example, that loans are evaluated by a group including individuals with biases (e.g., against immigrant or older loan applicants). If an algorithm is trained to mimic the decisions of these individuals (i.e., predict the loan approval decision) it will implicitly reproduce these biases. However, if the algorithm is trained to predict the future return of the investments made (i.e., the loans), the algorithm may show that unbiased decisions lead to better outcomes and may therefore provide better recommendations. In that case the AI algorithms might, for example, better align loan decisions with firms’ long-term profit interests while reducing biases [44].

AI methodologies are therefore not inherently good or bad, biased or clean. They are what we make of them. Interestingly, AI and claims data can be used to analyze the scope of the potential insurance discrimination problem, by examining the insurance premium distribution and identifying the groups that benefit for different scenarios of harmonization versus tailoring of premium rates.