Data collection could stump next phase of pre­dic­tive an­a­lyt­ics

Ag­gre­gat­ing claims and clin­i­cal data pose an ini­tial hur­dle for sys­tems seek­ing to use big data to tar­get pa­tients for pre­ven­tive in­ter­ven­tions

Modern Healthcare - - ACCOUNTABLE CARE - By Me­lanie Evans

Ad­vo­cate Health Care’s first foray into pre­dict­ing pa­tients’ fu­ture med­i­cal needs fo­cused on those at great­est risk for re­peat hos­pi­tal­iza­tions.

It made sense. Medi­care two years ago be­gan pe­nal­iz­ing hos­pi­tals with ex­ces­sive pa­tient read­mis­sions within 30 days of dis­charge. The pol­icy set hos­pi­tals scram­bling to iden­tify and head off po­ten­tial re­peat vis­i­tors. Penal­ties to date have cost hos­pi­tals more than $500 mil­lion, ac­cord­ing to the Ad­vi­sory Board, in­clud­ing as much as $5 mil­lion for some Ad­vo­cate hos­pi­tals.

Ad­vo­cate’s ini­tial in­vest­ment in pre­dic­tive an­a­lyt­ics paid off. The big-data ini­tia­tive, which com­bined in­for­ma­tion gleaned from pa­tients’ med­i­cal his­tory, claims, de­mo­graph­ics, lab­o­ra­tory re­sults, phar­macy use and pa­tients’ self- de­scrip­tion of their health sta­tus, was 20% more ac­cu­rate than al­ter­na­tive al­go­rithms in the mar­ket­place in pre­dict­ing who might be read­mit­ted af­ter dis­charge, sys­tem of­fi­cials said.

But now, with the new sys­tem in place at eight of its 11 hos­pi­tals, Ad­vo­cate is look­ing to take the strat­egy to the next level. The Down­ers Grove, Ill.based health sys­tem and its med­i­cal records part­ner later this year will launch a pre­dic­tive-an­a­lyt­ics ini­tia­tive that re­views all pa­tients re­ceiv­ing care from af­fil­i­ated physi­cians. The goal is to iden­tify pa­tients who are likely can­di­dates for in­ter­ven­tions to pre­vent dis­ease, bet­ter man­age their health con­di­tions out­side the hospi­tal and pre­vent fu­ture hos­pi­tal­iza­tions, all of which could save in­sur­ers and the sys­tem money.

The model sorts pa­tients by the com­plex­ity of their con­di­tions, and then iden­ti­fies those fac­tors that sig­nal those who are ripe tar­gets for in­ter­ven­tion such as un­filled pre­scrip­tions

or poor com­mu­ni­ca­tion be­tween pa­tients’ mul­ti­ple providers.

For providers, the pre­ven­tive in­ter­ven­tions en­abled by pre­dic­tive an­a­lyt­ics could deliver prof­its un­der new pay­ment mod­els, which are mov­ing to­ward var­i­ous forms of cap­i­ta­tion. And for pol­i­cy­mak­ers, the sav­ings could ease the fis­cal stress that U.S. health spend­ing puts on tax­pay­ers and em­ploy­ers.

“It’s en­abling strate­gic re­source al­lo­ca­tion among the to­tal pop­u­la­tion,” said Dr. Rishi Sikka, Ad­vo­cate’s vice pres­i­dent of clin­i­cal trans­for­ma­tion. “If you re­ally want to move the en­tire pop­u­la­tion … you need to work on the en­tire pop­u­la­tion, not just the most ex­pen­sive.”

Ad­vo­cate’s push to em­ploy pre­dic­tive mod­el­ing across its broad pop­u­la­tion base is an early test of the lat­est front in health­care’s march to us­ing big data to im­prove health­care out­comes and re­duce costs. It is oc­cur­ring against a back­drop where much of the in­dus­try is still strug­gling to boost the weak to mod­est ac­cu­racy of ex­ist­ing mod­els, which fo­cus on pre­vent­ing hospi­tal read­mis­sions and have so far yielded only mod­est re­sults (some no bet­ter than flip­ping a coin), ac­cord­ing to a 2011 re­view of more than two decades of stud­ies and more re­cent pub­lished re­search.

“Most mod­els were not very good at dis­crim­i­nat­ing be­tween pa­tients who were and were not go­ing to be read­mit­ted to the hospi­tal,” said Dr. De­van Kansagara, di­rec­tor of the Ev­i­dence­based Syn­the­sis Pro­gram at the Port- land Vet­er­ans Af­fairs Med­i­cal Cen­ter, as­sis­tant pro­fes­sor of medicine at Ore­gon Health & Sci­ence Univer­sity, and lead au­thor of the 2011 study.

Us­ing big data and ap­ply­ing pre­dic­tive an­a­lyt­ics have been hot topics among hospi­tal of­fi­cials and con­sul­tants for sev­eral years now. But many sys­tems that are mov­ing to im­ple­ment big-data an­a­lyt­ics face huge tech­no­log­i­cal hur­dles, in­clud­ing in­com­plete data re­sid­ing in mul­ti­ple elec­tronic health records that of­ten can­not com­mu­ni­cate with one an­other.

Ex­perts also note that some sys­tems’ ef­forts are fall­ing short be­cause they rely ex­clu­sively on claims data, which don’t in­clude the de­tails and nuance pro­vided by med­i­cal records. Those records are dif­fi­cult to ob­tain and even more dif­fi­cult to sift. One prob­lem: Mas­sive clin­i­cal data files con­tain loads of re­dun­dant in­for­ma­tion. “Prac­ti­cally, it’s just hard to ex­tract the data,” said Ian Dun­can, a con­sul­tant for the So­ci­ety of Ac­tu­ar­ies and an ad­junct pro­fes­sor of ac­tu­ar­ial sta­tis­tics at the Univer­sity of Cal­i­for­nia Santa Bar­bara.

Ad­vo­cate grap­pled with all those chal­lenges. Its deeper dive into pre­dic­tive an­a­lyt­ics re­quired roughly 18 months of work to merge, clean and or­ga­nize pa­tient data from mul­ti­ple sources, in­clud­ing ex­ter­nal in­sur­ance claims, in­ter­nal fi­nan­cial and de­mo­graphic records and mul­ti­ple elec­tronic med­i­cal-record sys­tems. The work was nec­es­sary, ex­ec­u­tives said, to deliver the com­mu­nity-wide health im­prove­ment—and sav­ings—that come from treat­ing fewer chron­i­cally ill pa­tients who show up on a hospi­tal’s doorstep with com­plex and hard-to-treat med­i­cal emer­gen­cies.

Ad­vo­cate de­vel­oped its new model with Cerner Corp. Af­ter scor­ing all pa­tients on mea­sures of well­ness and ill­ness com­plex­ity us­ing lab­o­ra­tory val­ues, pre­scrip­tion data, vi­tal signs and smok­ing sta­tus, the model al­lowed Ad­vo­cate’s providers to iden­tify those pa­tients whom doc­tors could ei­ther keep healthy or help avoid hos­pi­tal­iza­tion by plac­ing them in care-co­or­dina-

tion pro­grams, long-term care or home-care ser­vices. Pa­tients may also be iden­ti­fied who could ben­e­fit from less re­source-in­ten­sive in­ter­ven­tions, such as tele­phone out­reach.

The ACO im­per­a­tive

The boom­ing in­ter­est in pre­dic­tive mod­el­ing comes as hospi­tal rev­enue and mar­gins are in­creas­ingly de­pen­dent on providers’ skills at man­ag­ing care costs be­cause of the de­mands of ac­count­able care. Prof­its un­der ac­count­able-care con­tracts de­pend in part on providers’ suc­cess at keep­ing med­i­cal spend­ing be­low set tar­gets.

Ac­count­able-care con­tracts have more than tripled un­der Medi­care since 2012. To date, few have seen fi­nan­cial pay­outs un­der the model; three-quar­ters failed to earn bonuses dur­ing the first year. The pri­vate sec­tor’s in­ter­est has also blos­somed. There are now 626 ac­count­able care or­ga­ni­za­tions na­tion­wide, cov­er­ing more than 20 mil­lion in­di­vid­u­als, the con­sult­ing group Leav­itt Part­ners es­ti­mated in June.

As ac­count­able-care con­tracts cover more lives, providers will find them­selves at greater risk for steep losses if they fail to get costs un­der con­trol, since many are mov­ing to­ward cap­i­tated pay­ments.

The growth of ACOs has spurred a boom­ing con­sult­ing in­dus­try for data an­a­lyt­ics in health­care, with a rush of ven­dors jostling for busi­ness along­side some of the big­gest names in com­put­ers, soft­ware and con­sult­ing. Some ma­jor health­care de­liv­ery sys­tems in the coun­try are ei­ther con­sid­er­ing or in the early stages of im­ple­ment­ing pre­dic­tive-an­a­lyt­ics pro­grams.

This es­ca­lat­ing push into pre­dic­tive an­a­lyt­ics is ul­ti­mately be­ing driven by the grow­ing recog­ni­tion by some hos­pi­tals and med­i­cal groups that those who can iden­tify and pre­vent avoid­able trips to the emer­gency room, hospi­tal or clinic will be fi­nan­cially re­warded un­der ac­count­able care-like pay­ment mod­els. “If you’re more ef­fi­cient when you bear eco­nomic risk,

“If you re­ally want to move the en­tire pop­u­la­tion … you need to work on the en­tire pop­u­la­tion, not just the most ex­pen­sive.” DR. RISHI SIKKA, VICE PRES­I­DENT OF CLIN­I­CAL TRANS­FOR­MA­TION. AD­VO­CATE HEALTH CARE

that’s your profit mar­gin,” said Dr. David Nash, dean of the Thomas Jef­fer­son Univer­sity School of Pop­u­la­tion Health and a pro­fes­sor of health pol­icy.

Pre­dic­tive an­a­lyt­ics has been made pos­si­ble by the flood of data re­leased by health­care’s re­cent in­vest­ments in in­for­ma­tion tech­nol­ogy. To an­a­lyze the data trove, clin­i­cians in­creas­ingly work along­side statis­ti­cians, pro­gram­mers and ac­tu­ar­ies as they look for new op­por­tu­ni­ties to bet­ter man­age chron­i­cally ill pa­tients.

“Un­der­stand­ing as much as you can about all as­pects of your pa­tient, not just their dis­ease, but their so­cial set­ting, their his­tory of uti­liza­tion, their risk for hos­pi­tal­iza­tion, that’s big data in health­care,” Nash said. “The more you un­der­stand, the more ef­fi­ciently you can de­ploy re­sources.”

Pro­po­nents of pre­dic­tive an­a­lyt­ics say the new sys­tems will move be­yond the im­me­di­ate goal of re­duc­ing hospi­tal vis­its to fo­cus on preven­tion. They want to head off chronic dis­ease be­fore it starts or re­duce com­pli­ca­tions among those who are al­ready ill.

“All of the at­ten­tion up un­til now has been on the hos­pi­tals,” Port­land VA’s Kansagara said. “The next fron­tier is look­ing from the van­tage point of the pri­mary-care med­i­cal home. I think that some would ar­gue that we’re miss­ing the boat if we’re only look­ing at the hospi­tal.”

Pop­u­la­tion health 2.0

As hos­pi­tals and med­i­cal groups seek to cap­i­tal­ize on big data, they face sig­nif­i­cant in­for­ma­tion tech­nol­ogy chal­lenges, rang­ing from the ba­sic need for an elec­tronic health record to the in­ten­sive process of com­bin­ing data from mul­ti­ple sys­tems. Ad­vo­cate’s push into pre­dic­tive an­a­lyt­ics un­der­scores those chal­lenges.

Ad­vo­cate’s work with Cerner to bol­ster its an­a­lyt­ics ca­pa­bil­i­ties be­gan more than three years ago. Early ef­forts fo­cused on merg­ing and stan­dard­iz­ing clin­i­cal, claims and fi­nan­cial data across mul­ti­ple elec­tronic-health records. Ag­gre­gated data pro­vided the lon­gi­tu­di­nal data­base of pa­tient records for anal­y­sis, a project that re­quired 18 months to com­plete. “This work, in and of it­self, is ground­break­ing,” Sikka said.

The ef­fort re­quired data ex­trac­tion across mul­ti­ple data­bases with dif­fer­ent codes that needed to be stan­dard­ized, from pa­tient gen­der to lab­o­ra­tory val­ues. For ex­am­ple, one data­base might re­port gen­der with the word “fe­male,” an­other with the let­ter “f,” an­other with bi­nary code. Data­bases even used dif­fer­ent codes to iden­tify the same pa­tient, mak­ing even the most fun­da­men­tal work of ag­gre­gat­ing data a ma­jor hur­dle. Com­pre­hen­sive data “is the linch­pin,” said Tina Es­pos­ito, vice pres­i­dent for the Cen­ter for Health In­for­ma­tion at Ad­vo­cate.

Many providers lack the tech­nol­ogy or re­sources to over­come those bar­ri­ers, said Ariel Bayewitz, vice pres­i­dent of provider an­a­lyt­ics and reporting for ma­jor in­surer Well--

Point. Not all providers have EHRs and soft­ware varies among those that do. Wel­lPoint’s pre­dic­tive an­a­lyt­ics tools rely on claims data to iden­tify high-risk pa­tients by study­ing 40 con­di­tions or risk fac­tors, such as un­filled pre­scrip­tions.

Wel­lPoint is test­ing the in­te­gra­tion of health-sys­tem clin­i­cal data into pre­dic­tive mod­els through pi­lots with HealthCore, its re­search sub­sidiary. The com­pany de­clined to say how many sys­tems are in­volved in the ef­fort, but a spokes­woman said the work “so far has been fo­cused on the com­plex task of in­te­grat­ing clin­i­cal and claims data, and test­ing that data to en­sure we are ac­cu­rately cap­tur­ing the in­for­ma­tion to pro­vide the most ac­cu­rate lon­gi­tu­di­nal pa­tient record.” The data has been used to un­cover un­di­ag­nosed con­di­tions and track drug com­pli­ance, she said.

At Op­tum Labs, the Cam­bridge, Mass.-based col­lab­o­ra­tion be­tween health­care com­pa­nies and Op­tum, the con­sult­ing and an­a­lyt­ics arm of Unit­edHealth Group, re­searchers are test­ing use of pre­dic­tive mod­els for pa­tient risk along­side more re­tail- friendly an­a­lyt­ics com­monly used by Ama­zon or Net­flix to an­tic­i­pate con­sumer needs, said lab­o­ra­tory CEO Paul Ble­icher. “One of the rich­est op­por­tu­ni­ties, in terms of im­prov­ing pa­tient care and re­duc­ing the cost of health­care, which is the dual fo­cus of ACOs … is through pre­dic­tive mod­el­ing,” he said.

That’s a break from strate­gies to tar­get the costli­est pa­tients—the 5% that ac­count for 25% of spend­ing— that fail to make that dif­fer­en­ti­a­tion, Ad­vo­cate’s Es­pos­ito said. “That top of the pyramid is any­thing and ev­ery­thing.”

For that rea­son, tar­get­ing the most ex­pen­sive pa­tients is “the first gen­er­a­tion of pop­u­la­tion health,” said Ad­vo­cate’s Sikka. Us­ing more so­phis­ti­cated al­go­rithms to iso­late med­i­cal needs and risk across a full range of pa­tients is “pop­u­la­tion health 2.0,” he said.

Yet tar­get­ing the most ex­pen­sive pa­tients still has its pro­po­nents, par­tic­u­larly among ACOs in the early stages of us­ing pre­dic­tive an­a­lyt­ics. The strat­egy was made pop­u­lar by the Cam­den Coali­tion of Health­care Providers, which tar­geted in­ter­ven­tion to the most costly pa­tients in Cam­den, N.J.

In­ter­moun­tain Health­care, the Salt Lake City sys­tem with 22 hos­pi­tals and its own health plan, launched its an­a­lyt­ics pro­gram to de­ter­mine how to re­duce spend­ing among the 1% of its pa­tients who ac­counted for 24% of care ex­pen­di­tures be­tween 2008 and 2012. “We did not know a lot about them,” said Scott Pin­gree, In­ter­moun­tain’s di­rec­tor of strate­gic plan­ning and chair of high-cost pa­tients and hot-spot­ting.

What they dis­cov­ered was ex­ten­sive frag­men­ta­tion among the care­givers for these com­plex pa­tients who had on aver­age a dozen at­tend­ing physi­cians. Less than half had a pri­mary-care doc­tor, even though a fourth of pa­tients had at least three chronic dis­eases.

The sys­tem used that knowl­edge to open a re­fer­ral-only clinic in Salt Lake City for high-use, high-cost pa­tients. Pin­gree and his staff also be­gan to wade through the data to iden­tify clus­ters of pa­tients with sim­i­lar needs. “There’s al­ways more to learn from data,” he said.

Ac­count­able care or­ga­ni­za­tions in­volved with the Brook­ings In­sti­tu­tion’s ACO Learn­ing Net­work are in­creas­ingly in­ter­ested in pa­tients with a “ris­ing risk” of com­pli­ca­tions who would ben­e­fit from early in­ter­ven­tions. Un­til now, most ACOs tar­geted the top 5% of most ex­pen­sive pa­tients, said Dr. James Col­bert, a con­sul­tant for the net­work and a fac­ulty mem­ber at Har­vard Med­i­cal School.

Go­ing be­yond the high-cost strat­egy, a few sys­tems are ex­per­i­ment­ing with col­lect­ing data on so­cial de­ter­mi­nants of health that are not con­tained in ei­ther claims or elec­tronic med­i­cal records, such as hous­ing or ac­cess to food. Three mod­er­ately suc­cess­ful pre­dic­tors of re­peat hospi­tal vis­its among heart fail­ure pa­tients in­clude eco­nomic sta­tus, fre­quent ad­dress changes and co­caine use, re­searchers re­ported in the jour­nal Med­i­cal Care in 2010.

The idea is to get a full pic­ture of in­di­vid­u­als most at risk, Col­bert said. Un­for­tu­nately, most pre­dic­tive mod­els over­look so­cial de­ter­mi­nants of health, Kansagara and col­leagues re­ported in 2011.

Even the big­gest pro­po­nents of us­ing big data to con­duct pre­dic­tive an­a­lyt­ics do not think it will re­duce re­liance on the in­di­vid­ual physi­cian’s re­la­tion­ship with pa­tients. Many ACO ini­tia­tives in­clude the di­rect ex­pe­ri­ence of clin­i­cians with pa­tients—and their clin­i­cal in­tu­ition—to round out pa­tient por­traits of who may be at risk, said Dr. Farzad Mostashari, the for­mer na­tional co­or­di­na­tor for health in­for­ma­tion tech­nol­ogy, who is now chief ex­ec­u­tive of ACO con­sul­tancy Aledade.

Some large sys­tems are still hold­ing off, think­ing pre­dic­tive an­a­lyt­ics en­abled by big data may not be worth the in­vest­ment—at least not yet. Of­fi­cials at Ban­ner Health in Phoenix are mov­ing to a team ap­proach to de­velop pop­u­la­tion-health man­age­ment across all their ser­vices, in­clud­ing pre­ven­tive care and chronic-dis­ease man­age­ment, said Dr. Robert Groves, Ban­ner’s vice pres­i­dent for health man­age­ment.

Ban­ner, whose ACO formed in July 2011, cov­er­ing 285,000 Medi­care and commercial-con­tract ben­e­fi­cia­ries, is look­ing for ways to move be­yond tar­get­ing the most ex­pen­sive 5% of pa­tients to tar­get­ing the 15% of pa­tients most likely to end up in that most-ex­pen­sive group. But so far, it hasn’t found a pre­dic­tive-an­a­lyt­ics model that works any bet­ter than smart doc­tors. “I’m unim­pressed with their abil­ity,” Groves said. “It’s chess be­fore Big Blue.”

Ad­vo­cate Christ Med­i­cal Cen­ter, Oak Lawn, Ill., where of­fi­cials de­ployed a pre­dic­tive model us­ing clin­i­cal and claims data.

Newspapers in English

Newspapers from USA

© PressReader. All rights reserved.