Table 1. Uses of machine learning in health financing and their potential risks and benefits.
| Domain of use (function) | Type of application | Potential benefits | Risks |
|---|---|---|---|
| Prediction of health expenditure or health-care costs, also in relation to specific diseases or in relation to use by specific individuals or groups (revenue raising, purchasing) | Supervised learning (classification, variable selection, regression, prediction) | Through the revelation of hidden patterns in data (such as specific patient attributes or use patterns), more accurate cost predictions are possible (compared to traditional statistical methods), e.g. for specific diseases, for high-cost patients and for improved service tailoring (e.g. optimized disease management and prevention activities for specific population groups). These insights can contribute to better forecasting, more efficient spending, more equitable resource distribution and improved utilization in line with need. |
Predictive analytics could be used for the exclusion of high-cost diseases and introduction of cost-reduction measures at the expense of financial protection and quality of care. Predictive analytics endangers a person's privacy, i.e. a person may not give consent for predictions to be made based on their personal data. |
| Assessment of health risks in a pool (i.e. calculating patients’ health risk scores, a metric to predict aspects of a patient’s future care (e.g. comparing costs, risk of hospitalization of a patient with the average) (pooling) | Supervised learning (regression, variable selection) | More precise risk scoring through machine learning supported data analysis can improve risk adjustment or risk equalization mechanisms and formulas, which serve to ensure equalized per capita allocations or expenditure that reflects differences in health risks and needs across different pools. More precise scoring can contribute to more efficiency and equitable allocation of resources. | More precise risk scoring may facilitate risk selection and the exclusion of high (cost) users or otherwise vulnerable groups from health insurance, or the increase of their premiums (based on medical underwriting). These practices could lead to further fragmentation of risk pools and hence reduced equity in resource distribution, as well as increased inequity in access to care and reduced financial protection. |
| Claims review and fraud detection (purchasing) | Mostly supervised learning (classification), some examples of unsupervised learning (clustering/outlier detection) | By identifying patterns, machine learning-supported claims review can improve and accelerate: preauthorization of patient care; claims adjudication (the checking of validity and eligibility); proactive identification of coding and billing errors before claims processing and payment; and detection and investigation of outliers and duplicate claims. These improvements can result in reduced claims processing time, reduced human resources needed for claims review, reduced administrative costs, less erroneous payments and higher overall efficiency. Prudence in claims preparation by providers can be enhanced (less gaming), leading to higher transparency and accountability of health providers and purchasing agencies in their claims management activities. Possible synergy between automated (machine learning-supported) claims review and review by humans (whereby the former can support the latter) might lead to less subjectivity and errors (more transparency of purchasing agencies in their claims review decisions). |
Machine learning-supported claims management may enable more advanced, thus simplified, opportunities for surveillance of health service providers by the purchaser, potentially leading to a culture of fault-finding and mistrust among providers towards the purchaser. Machine learning-supported claims management could reduce the role of human judgment in claims review, which could lead to a lack of transparency and explicability of outcomes of purchasing agencies, and as a result of algorithmic bias, could lead to discrimination against certain population groups. |
| Design or revision of provider payments through claims analysis (purchasing) | No information on the applied types of machine learning | The use of machine learning results can provide more granular and more accurate insights to inform policy decisions on: provider payment methods and rates; and quality improvement, based on comparison of treatment practices across providers. Such policy decisions can result in improved efficiency, more equitable distribution of resources and more equitable access to care, financial protection and improved quality of care. |
Machine learning-supported data analysis may suffer from algorithmic bias and/or use of insufficient or unrepresentative data. Bias could lead to unintended, distorted outcomes, including unequal treatment of providers and potential market distortions. Lack of explicability of the working and outcomes of machine learning to health-care providers could erode trust. |
| Provider performance and contract monitoring by purchasers (purchasing) | No information on the applied types of machine learning | Automated performance monitoring and contract monitoring of providers by the purchaser can result in resource and time savings and elimination of human errors. | |
| Design of benefits and access conditions based on claims analysis (purchasing) | Unsupervised learning (clustering); also supervised learning (regression/ prediction) | Machine learning-based claims analysis can provide more granular insights on patient needs and expenditure trends to inform policy decisions on premiums and co-insurance rates (as cost determinants become clearer) and health needs-oriented benefits design. Using these insights for policy decisions can result in increased efficiency, more equitable distribution of resources and equitable access to care, quality of care and responsiveness to patient preferences and needs, and financial protection. |
Machine learning-supported data analysis may facilitate algorithmic bias and/or use of insufficient or unrepresentative data for policy decisions, potentially leading to unintended and distorted outcomes, including discrimination against individuals or population groups and increasing inequities in resource distribution. Lack of explicability of the working and outcomes of machine learning to beneficiaries could erode trust of people in social protection schemes and programmes. Combination of data from multiple databases can increase opportunities for surveillance and lack of privacy. |
| Identification of beneficiaries for targeting policies (pooling and purchasing) | Supervised learning (classification, variable selection, regression); unsupervised learning (clustering) | Precision and efficiency can be increased in targeting processes for poor and/or vulnerable beneficiaries of government-funded health coverage or cash transfer programmes, as well as in setting differentiated and tailored cost-sharing policies, for example by applying machine learning to multiple, combined data sets. Such improved precision can lead to higher administrative efficiency and more equitable financial protection. |
|
| Beneficiary enrolment (pooling) | Supervised learning (classification) | Identity confirmation, identification of beneficiaries and timely enrolment can be improved, which can reduce coverage gaps (exclusion errors) and increase financial protection. Validation of data entry in beneficiary databases, e.g. through identity confirmation and authentication support processes, improves data quality. As a result, fraud, abuse, inclusion errors and unnecessary data collection can be reduced. |