HealthKonnect, a subsidiary of Chinese insurer Ping An Insurance Group developed a personal health-risk prediction model that has the ability to predict at individual patient level future healthcare risks measured by total medical costs. Here’s how.
By combining techniques from traditional health actuarial risk-prediction modeling and big data machine learning, HealthKonnect, a managed care subsidiary company under Chinese insurer Ping An Insurance Group, has developed a personal health-risk prediction model that can predict future healthcare risks measured by total medical costs for individual patients.
Established in 1988, Ping An, headquartered in Shenzhen, China, is the first integrated financial services conglomerate in China that blends its core insurance operations into securities brokerage, trust and investment, commercial banking, asset management, and corporate pension business.
“So far, our predictive model and big data analytics have been applied in an outpatient chronic disease management pilot in a fourth-tier city in China,” says Zheng Yi, chief medical officer and senior vice president at HealthKonnect. “The local Social Health Insurance (SHI) office has experienced over 20% medical trend in the past three years. After we took over the management responsibility of this population, we used predictive modeling and a big data analytics-driven management tools and effectively bucked the trend to -1.2% within nine months.”
In addition, HealthKonnect has built a big data-based fraud, waste, and abuse (FWA) model to improve the ability to manage FWA of medical resources in healthcare system.
“These series of models are being utilized in population management of a chronic disease population in China,” says Zheng. “In 2017, our health analytic-driven interventions resulted in significant cost reduction and brought the annual cost trend from 23% down to -1%.”
Here, Zheng discusses HealthKonnect’s strategy and model, as well as current results.
Managed Healthcare Executive (MHE): How did this idea begin?
Zheng: SHI, which is a government-funded program, pays for 48% of total medical expenditure in China. Current benefit administration managed through the local SHI office is built on fee-for-service payment scheme. There’s very limited management capability from system and data analytics within the SHI administration. Escalating healthcare spend presents a huge challenge to SHI administrators from both budgeting and fund management perspective. Historically, SHI funding was determined by prior year spend and capitation methods. Furthermore, FWA, though a known issue to SHI, is extremely difficult to detect. Rule-based utilization management program has had limited success due to high rate of false positives. Predictive modeling can help accurately forecast future medical spend and reduce variability in the budgeting process. In addition, big data analytics provide a new and exciting way that expands beyond prior knowledge dependent rules engine to detect FWA through a multi-dimension machine learning approach.
MHE: Why was it implemented? What are the current results?
Zheng: Historically the FWA program is built on benefit policy and clinical knowledge-based rules. These are often applied through software that’s embedded in hospital system or ad hoc retrospective claims analysis provided by third-party vendors. The result has been very limited due to high rate of false positives that requires labor intensive manual review by claim investigation specialists contracted by the government. Our big data FWA model was developed to improve effectiveness of our rule based program while reduce cost of manual review. Preliminary result indicates big data approach can improve effectiveness of FWA detection by five to ten times while reducing false positives up to 75%.
MHE: Where are you getting the data to determine individual medical expense? How do you deal with the fact that costs vary by provider, region, and other factors?
Zheng: HealthKonnect provides FWA services and actuarial services to many local and regional SHI offices. Analytic insights acquired through our service were used to develop our big data analytic models. While there are a lot of variations by providers and local policies, we leverage our proprietary grouper and health risk adjustment model to account for variations in data. Big data techniques such as K-means and t-SNE were used to improve modeling capability to deal with variations. In addition, we use both supervised and unsupervised machine learning technique to enhance our modeling performance against local and regional variations.
MHE: How are you impacting fraud and waste?
Zheng: Our big data analytics serves as a foundation to support fraud waste and abuse management. Output of our model is built in management tools that are used by FWA claim investigators for targeted provider, pharmacy, and patient interventions. Once a fraud pattern is identified and confirmed by our big data model, it can also be translated into rules that are updated in our rule based engine embedded in hospital monitoring system for concurrent utilization review. Confirmed fraudulent activities also result in payment recovery from providers and disciplinary actions against eligibility of providers and patients.
MHE: How replicable would this be in the United States?
Zheng: While the healthcare system is very different between China and the U.S., leveraging the recent advancement in machine learning and big data analytics is applicable in healthcare management regardless of systems. Both countries face escalating healthcare cost trends and challenges of fighting ever-changing FWA patterns, that is costing the system a large amount of money. We believe building an effective predictive model and a big data antifraud model are essential in managing healthcare spend. Our experience has shown such advance analytics can help close some of the current issues within the traditional rule-based utilization management program, improve effectiveness while reduce operational cost. In addition, it helps identify new cost savings opportunities that are hard to identify solely based on prior knowledge and experience.