ABSTRACT

Objectives: To decide whether or not a threat prediction mannequin utilizing synthetic intelligence (AI) to mix a number of knowledge sources, together with claims knowledge, demographics, social determinants of well being (SDOH) knowledge, and admission, discharge, and switch (ADT) alerts, extra precisely identifies high-cost members than conventional fashions.

Study Design: The research used knowledge from a Medicaid accountable care group and included a inhabitants of 61,850 members repeatedly enrolled between May 2018 and April 2019.

Methods: Risk scores generated by 2 fashions had been estimated for every member. One mannequin, developed by Medical Home Network, used AI to investigate SDOH knowledge, ADT exercise, and claims and demographic traits, whereas the opposite mannequin (Chronic Illness and Disability Payment System [CDPS]) relied solely on demographic and claims data. To evaluate fashions, we calculated imply, median, and complete spending for members with the best 5% of AI threat scores and in contrast these with spending metrics for members with the best 5% of CDPS scores. We additionally in contrast the variety of members with the best 5% of prices prospectively recognized by every mannequin as highest threat. We segmented the inhabitants by size of prior enrollment to manage for various ranges of claims expertise.

Results: The AI mannequin constantly recognized a better proportion of the highest-spending members. Members deemed highest threat by the AI mannequin additionally had increased spending than members deemed highest threat by the CDPS mannequin.

Conclusions: Identification of high-cost members could be improved through the use of AI to mix conventional sources of information (eg, claims and demographic data) with nontraditional sources (eg, SDOH, admission alerts).

Am J Manag Care. 2022;28(11):In Press

_____

Takeaway Points

Prediction fashions that complement claims data with knowledge on social determinants of well being, care administration data, and admission, discharge, and switch alerts and analyze these knowledge utilizing synthetic intelligence (AI) can determine members with excessive future spending extra precisely than regression-based fashions that rely solely on claims and demographic knowledge.

• AI permits threat stratification packages to combine a number of knowledge sources right into a single, actionable mannequin.
• More well timed, dynamic, and correct threat stratification allows allocation of care administration sources to sufferers on whom care administration can have the best impression.
• Care administration packages ought to gather and use data aside from claims knowledge for threat stratification.

_____

Attempts to decrease medical prices by figuring out and intervening within the care of sufferers with the best want, generally known as “hotspotting,” is a theoretically enticing method to decreasing well being care spending and enhancing inhabitants well being.¹ Unfortunately, a current analysis of some of the distinguished hotspotting packages discovered that this system failed to cut back readmissions and prices.² Evaluation outcomes counsel that the optimistic results attributed to related care administration packages might consequence from analysis designs that don’t adequately management for regression to the imply, slightly than the power of those care administration packages to cut back prices. This highlights a problem that many care administration packages face: Although it’s straightforward to determine sufferers with excessive previous utilization, these sufferers won’t essentially be the best utilizers sooner or later.^3-5

Prior utilization data is more likely to be an vital component in figuring out high-cost sufferers, however ideally it will be mixed with different knowledge to determine the rising threat inhabitants. Collecting doubtlessly helpful knowledge corresponding to admission, discharge, and switch (ADT) feeds from hospitals and integrating them into predictive modeling are problematic in apply. As a consequence, well being plans and state Medicaid companies have used threat prediction fashions based on claims knowledge.^6,7 However, these claims-based actuarial fashions lack data on the social determinants of well being (SDOH) related to medical spending.^8,9 This is the results of difficulties in accumulating and integrating member-specific SDOH knowledge into processes for threat stratification and directing care administration sources, which have proved difficult for insurers and suppliers alike. Although many well being insurers acknowledge the necessity to tackle SDOH and have been lively in creating packages to take action,¹⁰ creating the infrastructure wanted to gather individual-level SDOH knowledge has been more difficult. Even accountable care organizations (ACOs) actively engaged in packages to deal with SDOH battle to evaluate members’ SDOH wants and to include this data into focused interventions.¹¹ Providers have additionally reported related challenges in accumulating and utilizing SDOH knowledge.¹²

In addition to missing knowledge on SDOH elements affecting utilization and price, claims-based fashions are topic to lags between service provision and claims receipt. This leads to an absence of real-time details about inpatient admissions, discharges, and transfers, in addition to emergency division (ED) visits. The capacity to quickly reply to those acute medical occasions is an integral part of profitable care administration packages,¹³ and this response is hindered when ADT data shouldn’t be included in threat stratification fashions.

Care administration packages are restricted not solely by the information to which they’ve entry but in addition by the fashions used to leverage these knowledge. Many fashions fail to reap the benefits of current developments within the discipline of synthetic intelligence (AI), relying as a substitute on primary regression methods or comparatively easy algorithms primarily based on prior utilization of particular companies.

In this text, we study the efficiency of a threat prediction rating developed by Medical Home Network (MHN) that’s primarily based on each conventional knowledge parts accessible to managed care organizations (together with demographic data and medical/pharmacy claims) and nontraditional knowledge parts (together with SDOH, ADT, and care administration data). This threat rating is constructed utilizing AI. We evaluate its efficiency in opposition to the Chronic Illness and Disability Payment System (CDPS) threat rating, which predicts future spend utilizing solely conventional knowledge parts.

DATA AND METHODS

Study knowledge had been offered by an ACO delegated for care administration of a Medicaid inhabitants in Cook County, Illinois. This inhabitants consists of people eligible for Medicaid by way of the Affordable Care Act’s Medicaid enlargement provision, Medicaid-eligible moms and kids (known as the “family health plan”), and people eligible for Medicaid due to incapacity (the “integrated care program”).

For this research, we analyzed spending within the 12 months between May 2018 and April 2019. We excluded people with any pregnancy-related spending in the course of the research interval as a result of the chance prediction fashions that we examined usually are not calibrated to foretell pregnancy-related spending and since pregnant members are referred to care administration by means aside from the AI mannequin and are served by a special care administration program. This resulted in exclusion of 7480 members from the pattern. In addition, we excluded 11,446 people engaged in case administration in some unspecified time in the future throughout their enrollment within the ACO. We selected to exclude these members as a result of case administration efforts might have an effect on the spending that our threat scoring fashions tried to foretell. Finally, we excluded 76,911 members not repeatedly enrolled in the course of the research interval and 221 members with lacking threat rating data. Our last pattern consisted of 61,850 people repeatedly enrolled over 12 months. Table 1 gives particulars on how our pattern was derived.

Risk Models

We in contrast 2 fashions used to evaluate threat for members. The first, CDPS, is designed to foretell threat for Medicaid populations¹⁴ and is utilized by many state Medicaid companies to set charges for Medicaid managed care plans. CDPS is a regression-based mannequin that depends on medical and pharmacy claims historical past to foretell spending primarily based on diagnoses and demographic elements. The CDPS mannequin consists of clusters of diagnoses outlined by a mix of empirical evaluation and medical judgment.

The second threat mannequin, referred to right here because the AI mannequin, makes use of the identical demographic variables (together with member age) and medical and pharmacy claims knowledge because the CDPS mannequin however provides a number of further knowledge sources, together with SDOH knowledge, ADT knowledge, and knowledge offered by care managers. SDOH data is collected utilizing a proprietary member survey referred to as the Health Risk Assessment (HRA). The HRA collects data on the most typical power sicknesses with potential for care administration impression, current inpatient or ED utilizations, and SDOH-related limitations to remedy adherence. The data collected by the HRA is listed within the eAppendix (available at ajmc.com). Surveys are administered to most members (roughly 85%) inside 60 days of plan enrollment and repeated primarily based on threat degree or on triggered occasions, corresponding to member request or a sudden improve in utilization. The HRA knowledge are fed into the AI mannequin, together with medical claims, pharmacy claims, care administration, and different administrative knowledge, to develop a person AI threat rating for every affected person. The AI mannequin then employs a machine studying regression mannequin to foretell the entire price of members’ medical claims. The coaching knowledge for the mannequin include historic member knowledge utilizing an 80/20 cut up to outline coaching and testing samples. Once a great set of hyperparameters is recognized, all the coaching set is used to create the ultimate model of the mannequin and examined on the withheld testing pattern.

We calculated predictions from every threat mannequin utilizing knowledge accessible as of May 2018, the start of the research interval, earlier than price measurement started. It is feasible that the fashions’ relative performances might differ for people with totally different lengths of claims expertise. Because the CDPS mannequin is primarily claims primarily based, it could be much less ready than the AI mannequin to determine pricey beneficiaries with out an in depth claims historical past, because the AI mannequin consists of each claims and nonclaims knowledge. However, the inclusion of nonclaims knowledge could also be much less vital as a member’s claims historical past grows. To account for this risk, we stratified our pattern into 3 totally different classes primarily based on prior claims expertise as of the start of the research interval. We estimated mannequin efficiency individually for members with 0 to three months of prior enrollment, members with 4 to 12 months of prior enrollment, and members with greater than 12 months of prior enrollment.

Analysis and Outcome Measures

Our aim was to evaluate how profitable every mannequin is at figuring out members with the highest 5% of spending (though we different the 5% definition of “high-spending” in sensitivity analyses as a result of totally different care administration organizations’ definitions of high-spending members focused for intervention might differ primarily based on the care administration sources a corporation has accessible). Spending was outlined as complete medical and pharmacy spending. We selected to look at the fashions’ capacity to determine members within the high 5% of spending slightly than to precisely estimate the precise greenback quantity of a member’s spending as a result of figuring out high-spending members is a extra vital step in allocating care administration sources than figuring out actual member spending quantities. Using strategies much like these of prior research,¹⁵ we in contrast threat prediction fashions by calculating imply, median, and complete spending for members with the best 5% of threat scores at the start of the research interval. We then in contrast these quantities in opposition to precise spending for members with the best 5% of prices from May 2018 to April 2019. Risk prediction fashions that efficiently determine high-cost members will determine extra complete and median spending. As an extra measure of every mannequin’s efficiency, we discovered the proportion of members within the highest 5% of spending who had been additionally within the highest 5% of threat scores.

RESULTS

Sample Characteristics

Table 2 reveals the traits of our inhabitants. The imply CDPS threat rating was 0.952, with a spread of 0.072 to 41.77. Similarly, the imply AI threat rating was 42.07, starting from 0 to 100. (The AI threat rating relies on predictive prices for all members remodeled to a 0-100 ranked scale.) Our pattern consists of 29,119 members with greater than 12 months of prior enrollment at the beginning of the research interval and 30,970 members with between 4 and 12 months of prior enrollment. Only 1761 members within the group had fewer than 4 months of prior enrollment. Consistent with the Medicaid inhabitants on the whole, our pattern is comparatively younger, with a imply age of 21.9 years, and is 56% feminine. Most members (79.1%) are Medicaid-eligible mother and father and kids, 15.7% of members are from the Affordable Care Act enlargement inhabitants, and 5.2% are Medicaid eligible due to a incapacity. Mean annual spending was $2070 per member, though spending variance was excessive, with an SD of $9904.

Spending Comparisons

Table 3 compares spending measures for the 5% of members with the best prices with these for the teams of members with (1) the best 5% of AI threat scores and (2) the best 5% of CDPS threat scores. Results are stratified by the variety of months that members had been enrolled earlier than the predictive threat scores had been computed. For members with the longest claims historical past (these with greater than 12 months of prior enrollment), the highest 5% of members by spending (n = 1456 members) incurred a complete of $38,543,492 in prices (imply, $26,472; median, $16,078). Members with the best 5% of AI threat scores incurred a complete price of $20,892,684 in the course of the research interval (imply, $14,349 per member; median, $7265). Forty-one % of members with the best AI threat scores had been additionally within the group of members with the best spend. By manner of comparability, imply and median prices for the group of members with the best 5% of CDPS threat scores had been decrease ($11,808 and $3753, respectively), and solely 29% of members with the best CDPS scores had been within the highest spending group. The AI mannequin’s high-risk group due to this fact had increased complete and imply and median spending than the CDPS mannequin’s high-risk group, and the AI mannequin recognized a bigger variety of the highest-cost members than the CDPS mannequin did.

Furthermore, the AI mannequin recognized 41% extra of the best spending members (175 further members) than the CDPS mannequin, and these members had $3.7 million extra in complete spending ($20,892,684 in contrast with $17,192,192). As a consequence, assigning care managers primarily based on AI threat scores provides managed care organizations the possibility to focus care administration sources on a better portion of high-cost members than in the event that they threat stratified utilizing CDPS scores.

It was initially anticipated that the AI mannequin would possibly outperform the CDPS mannequin among the many inhabitants of members with the fewest months of prior enrollment (and therefore the least claims historical past). However, as proven in Table 3, AI recognized extra members with the best 5% of precise spending and recognized members with increased imply, median, and complete spending than did CDPS, throughout all prior enrollment teams.

Sensitivity Analyses

Sensitivity analyses examined whether or not outcomes had been strong to our definition of high-cost members (ie, members with the best 5% of spending), our exclusion of case-managed enrollees, and give attention to complete spending versus medical spending alone. Primary outcomes outlined the high-cost group as these with the best 5% of spending, and we different this definition of high-cost, setting it on the high 1% after which the highest 3% of members by price. Similarly, sensitivity analyses had been performed together with the 11,446 members who participated in case administration. These analyses had been repeated utilizing medical spending as an final result (versus medical and pharmacy spending). In all instances, the outcomes had been much like these introduced above. These outcomes can be found from the authors upon request.

DISCUSSION

Many payers and supply methods interact in care administration efforts to cut back medical spending by figuring out sufferers more likely to incur excessive prices, then intervening to cut back preventable spending. Unfortunately, regardless of many accessible predictive fashions, identification of members with the best future spending stays difficult. Our outcomes counsel that that is due partly to the heavy reliance of those fashions on demographic and claims knowledge and their incapability to include different sources of information.

Identifying preventable spending might require figuring out sufferers with quickly rising threat scores, not simply sufferers whose threat scores are already excessive. In truth, the ACO studied is already concentrating on members with quickly rising threat scores for care administration. Understanding how you can finest incorporate threat rating adjustments into threat stratification efforts is a chance for future analysis. However, to the extent that this sort of threat trajectory evaluation can enhance threat stratification, plans with out the infrastructure to mix knowledge sources in actual time will miss care administration alternatives.

Some care administration organizations battle to assemble knowledge infrastructure and to create the processes which are obligatory to gather nonclaims knowledge, analyze knowledge from a number of sources, and use these to deploy care administration sources. This is no surprise, as all elements of this course of are difficult. Collecting SDOH data shouldn’t be a simple course of, and the business has not adopted normal devices for accumulating these knowledge. Unlike SDOH data, ADT data is on the market in a standardized format, however communities usually lack infrastructure to speak this data amongst hospitals, insurers, and care administration packages. Although stakeholders in some areas of the nation have collaborated to ascertain well being data exchanges to facilitate transmission of affected person knowledge amongst unaffiliated suppliers, these exchanges are nonetheless the exception slightly than the rule. As a consequence, many care administration packages should depend on claims knowledge which are usually no less than 3 months outdated because of the lag between service supply and claims processing. Finally, even organizations which have managed to gather and retailer knowledge from a number of disparate sources would possibly lack the experience to investigate the information in a manner that informs care administration efforts. Alternatively, some plans would possibly be capable of mix these knowledge however nonetheless depend on older analytical methods that lack the predictive enhancements that newer AI methods supply. Even plans that efficiently incorporate a number of knowledge units into AI fashions might face challenges making these knowledge usable to care managers. The group offering knowledge for this research identifies elements contributing to a member’s excessive threat rating in order that case managers can higher perceive the traits that make a member excessive threat.

Ultimately, a care administration program’s capacity to have an effect on utilization and price will rely on the interventions this system makes. However, even the very best interventions are unlikely to achieve success if they’re focused towards low-cost sufferers with little potential profit. These outcomes underscore the significance of care administration packages’ funding in improved knowledge infrastructure and evaluation.

Limitations

Our analyses are restricted in a number of methods. First, our knowledge come from a single ACO working in a single geographic space and serving a Medicaid inhabitants extra more likely to face SDOH-related limitations than members of a commercially insured inhabitants. For occasion, of the SDOH elements included within the AI mannequin, a number of the most extremely correlated with price had been wants associated to meals, clothes, or housing and self-reported well being ranking of honest or poor.⁸ These outcomes might not generalize to different ACOs or well being plans whose member populations differ considerably. However, even if that is so, many care administration packages serve related populations to whom these outcomes are more likely to generalize. Second, our knowledge are drawn from a single 12-month interval, May 2018 by way of April 2019. We can’t determine any distinctive occasions occurring throughout this era that might make the AI mannequin’s predictions extra correct than they’d be throughout different time intervals; nonetheless, it stays attainable that outcomes from different time intervals might differ. Third, though the AI mannequin was comparatively profitable at figuring out high-cost people, it was nonetheless unable to determine roughly half of the high-cost members. Part of this problem is pushed by the random nature of some well being spending. However, MHN is at the moment working so as to add different knowledge to the mannequin, notably medical laboratory knowledge, to enhance prediction. Finally, our analyses give attention to figuring out practices that may enhance the power to determine members who will incur excessive prices. Identification is step one in creating packages to manage medical spending, however the capacity to determine high-cost members doesn’t guarantee price reductions. Ultimately, the impact on spending will depend on the effectiveness of the care administration program, and profitable price discount is not at all assured. Several distinguished care administration packages have did not exhibit adjustments in utilization or price outcomes,^2,16 though others report larger success.^13,17,18 To tackle this subject, the care supervisor for the ACO being studied is creating disease-specific interventions to focus on circumstances which are each extremely prevalent and are related to excessive prices within the AI mannequin.

CONCLUSIONS

Many care administration packages have restricted sources of information on their members, and plenty of use regression-based strategies to determine members vulnerable to excessive spending. Our outcomes counsel {that a} mannequin developed utilizing AI and analyzing knowledge that embrace claims, demographics, SDOH, and ADT data can extra efficiently determine high-cost members than a mannequin primarily based on claims and demographic knowledge alone. We counsel that care administration packages can higher goal their interventions by investing within the infrastructure obligatory to gather, retailer, and replace new knowledge sources and within the experience to mix these knowledge utilizing superior analytic strategies.

Author Affiliations: College of Health Professions, Virginia Commonwealth University (NWC), Richmond, VA; Medical Home Network (AJ, TB, CL, KS, TP), Chicago, IL.

Source of Funding: None.

Author Disclosures: Dr Jones is a board member and worker of the nonprofit Medical Home Network. Mr Burkard is an worker of Medical Home Network. Ms Lulias is a board member of MoreCare and Medical Home Network, has consulted for MoreCare, is an worker of Medical Home Network, and owns inventory in MoreCare. Ms Posa is an worker of Medical Home Network. Dr Carroll and Ms Severson report no relationship or monetary curiosity with any entity that might pose a battle of curiosity with the subject material of this text.

Authorship Information: Concept and design (NWC, AJ, TB, CL, TP); acquisition of information (NWC, TB, CL, KS); evaluation and interpretation of information (NWC, AJ, TB, KS, TP); drafting of the manuscript (NWC, TB, CL, TP); crucial revision of the manuscript for vital mental content material (NWC, TP); statistical evaluation (NWC, TB, KS); acquiring funding (CL); administrative, technical, or logistic help (TB, CL, TP); and supervision (AJ, TB, CL, TP).

Address Correspondence to: Nathan W. Carroll, PhD, College of Health Professions, Virginia Commonwealth University, 900 E Leigh St, Richmond, VA 23298. Email: carrolln@vcu.edu.

REFERENCES

1. Gawande A. The scorching spotters. The New Yorker. January 16, 2011. Accessed January 12, 2021. https://www.newyorker.com/magazine/2011/01/24/the-hot-spotters

2. Finkelstein A, Zhou A, Taubman S, Doyle J. Health care hotspotting—a randomized, managed trial. N Engl J Med. 2020;382(2):152-162. doi:10.1056/NEJMsa1906848

3. Figueroa JF, Lyon Z, Zhou X, Grabowski DC, Jha AK. Persistence and drivers of high-cost standing amongst dual-eligible Medicare and Medicaid beneficiaries: an observational research. Ann Intern Med. 2018;169(8):528-534. doi:10.7326/M18-0085

4. Figueroa JF, Zhou X, Jha AK. Characteristics and spending patterns of persistently high-cost Medicare sufferers. Health Aff (Millwood). 2019;38(1):107-114. doi:10.1377/hlthaff.2018.05160

5. Johnson TL, Rinehart DJ, Durfee J, et al. For many sufferers who use massive quantities of well being care companies, the necessity is intense but momentary. Health Aff (Millwood). 2015;34(8):1312-1319. doi:10.1377/hlthaff.2014.1186

6. Hileman G, Steele S. Accuracy of Claims-Based Risk Scoring Models. Society of Actuaries; 2016. Accessed November 12, 2019. https://www.soa.org/globalassets/assets/Files/Research/research-2016-accuracy-claims-based-risk-scoring-models.pdf

7. Courtot B, Coughlin TA, Lawton EA. Medicaid and CHIP managed care fee strategies and spending in 20 states. Office of the Assistant Secretary for Planning and Evaluation. December 2012. Accessed November 12, 2019. https://aspe.hhs.gov/sites/default/files/migrated_legacy_files//43966/rpt.pdf

8. Jones A, Lemak CH, Lulias C, Burkard T, McDowell B, Severson Ok. Predictive worth of screening for addressable social threat elements. HSOA J Community Med Public Health Care. 2017;4(30). doi:10.24966/CMPH-1978/100030

9. Chen S, Bergman D, Miller Ok, Kavanagh A, Frownfelter J, Showalter J. Using utilized machine studying to foretell healthcare utilization primarily based on socioeconomic determinants of care. Am J Manag Care. 2020;26(1):26-31. doi:10.37765/ajmc.2020.42142

10. Berry Ok. How medical insurance suppliers are tackling social limitations to well being. Am J Accountable Care. 2019;7(4):19-21.

11. Murray GF, Rodriguez HP, Lewis VA. Upstream with a small paddle: how ACOs are working in opposition to the present to satisfy sufferers’ social wants. Health Aff (Millwood). 2020;39(2):199-206. doi:10.1377/hlthaff.2019.01266

12. Fraze TK, Brewster AL, Lewis VA, Beidler LB, Murray GF, Colla CH. Prevalence of screening for meals insecurity, housing instability, utility wants, transportation wants, and interpersonal violence by US doctor practices and hospitals. JAMA Netw Open. 2019;2(9):e1911514. doi:10.1001/jamanetworkopen.2019.11514

13. Hong CS, Siegel AL, Ferris TG. Caring for high-need, high-cost sufferers: what makes for a profitable care administration program? The Commonwealth Fund. August 7, 2014. Accessed November 12, 2019.
https://www.commonwealthfund.org/publications/issue-briefs/2014/aug/caring-high-need-high-cost-patients-what-makes-successful-care

14. Kronick R, Gilmer T, Dreyfus T, Lee L. Improving health-based fee for Medicaid beneficiaries: CDPS. Health Care Financ Rev. 2000;21(3):29-64.

15. Robst J. Comparing strategies for figuring out future high-cost psychological well being instances in Medicaid. Value Health. 2012;15(1):198-203. doi:10.1016/j.jval.2011.08.007

16. Peikes D, Chen A, Schore J, Brown R. Effects of care coordination on hospitalization, high quality of care, and well being care expenditures amongst Medicare beneficiaries: 15 randomized trials. JAMA. 2009;301(6):603-618. doi:10.1001/jama.2009.126

17. Kumar GS, Klein R. Effectiveness of case administration methods in decreasing emergency division visits in frequent person affected person populations: a scientific assessment. J Emerg Med. 2013;44(3):717-729. doi:10.1016/j.jemermed.2012.08.035

18. Holahan J, Schoen C, McMorrow S. The potential financial savings from enhanced power care administration insurance policies. Urban Institute. November 30, 2011. Accessed November 12, 2019. https://www.urban.org/sites/default/files/alfresco/publication-pdfs/412453-The-Potential-Savings-from-Enhanced-Chronic-Care-Management-Policies.pdf