Pre-empt­ing Care

Provider or­ga­ni­za­tions are us­ing com­put­ing power to de­ter­mine which pa­tients need help—be­fore crises oc­cur.

Health Data Management - - CONTENTS - BY GI­ENNA SHAW

An­a­lyt­ics al­low providers to iden­tify— and tighten—gaps in care for at-risk pa­tients.

Any­one who’s shopped on the ecom­merce jug­ger­naut Ama­zon knows that the com­pany an­a­lyzes cus­tomers’ shop­ping habits and serves up sug­ges­tions for what they might like to buy next. If you’re look­ing at one type of prod­uct, the site can tell you how many cus­tomers bought a dif­fer­ent item or what they bought in ad­di­tion to that item.

“Why can’t we do that in health­care?” won­ders Charles “Chuck” Chris­tian, vice pres­i­dent of tech­nol­ogy and en­gage­ment at the In­di­ana Health In­for­ma­tion Ex­change (IHIE).

Or­ga­ni­za­tions are “learn­ing now how to stan­dard­ize [data] and nor­mal­ize it us­ing good stan­dards and be­ing able to use that data to im­prove the qual­ity of care and the qual­ity of the out­comes,” Chris­tian says.

The next step is to serve it up to physi­cians be­fore they even need it to get pa­tients the right care at the right time and to use pre­dic­tive an­a­lyt­ics to keep pop­u­la­tions of pa­tients healthy, he con­tends.

If we can pre­dict shop­ping pref­er­ences, we should be able to put that tech­nol­ogy to work to pre­dict when some­one is at risk for read­mis­sion or might miss an ap­point­ment, adds Deb­o­rah Vi­ola, vice pres­i­dent of data man­age­ment and an­a­lyt­ics at Westch­ester Med­i­cal Cen­ter Health Net­work, a nine-hos­pi­tal sys­tem based in Val­halla, N.Y.

“That’s what makes this work ex­cit­ing, be­cause it has so much po­ten­tial,” she says.

Around the coun­try, health­care or­ga­ni­za­tions are tak­ing that next

step—us­ing data to pre­dict, in­ter­vene and pre­vent health crises. The goal of such ini­tia­tives is to serve at-risk pa­tients across the care con­tin­uum by re­duc­ing or elim­i­nat­ing gaps in care, to pre­vent read­mis­sions or costly and un­nec­es­sary emer­gency de­part­ment vis­its, and to im­prove pop­u­la­tion health with a fo­cus on pre­ven­tion for those who are most in need of tar­geted in­ter­ven­tion.

Ma­chine learn­ing, ar­ti­fi­cial in­tel­li­gence and pre­dic­tive an­a­lyt­ics are help­ing health­care in this quest.

Open­ing the tool­kit

Geisinger Health Sys­tem, for ex­am­ple, uses a va­ri­ety of tools to power its data an­a­lyt­ics pro­gram. The 13-hos­pi­tal in­te­grated sys­tem, based in Danville, Pa., has been an­a­lyz­ing his­tor­i­cal records gath­ered over more than a decade us­ing ma­chine learn­ing and ar­ti­fi­cial in­tel­li­gence.

Geisinger uses two big data plat­forms: the open-source Apache Hadoop and Cerner’s HealtheIn­tent plat­form. About 87 per­cent of its data is not dis­crete—it must be mined from the text in sources such as physi­cian notes.

“We use nat­u­ral lan­guage pro­cess­ing to scan through those notes, look­ing for spe­cific el­e­ments that we can ex­tract and use for an­a­lyt­ics pur­poses,” says CIO John Kravitz.

Ra­di­ol­o­gists’ notes, for ex­am­ple, are rich with data about imag­ing re­sults, such as the size of the aorta and de­vi­a­tions of the aor­tic wall. “We look at those notes again and com­pile that into a database. And if we see, for any rea­son, that the nod­ule is grow­ing in size, that trig­gers events from pre­dic­tive an­a­lyt­ics to iden­tify to the sur­geons that this pa­tient will most likely have an aor­tic aneurysm in the near fu­ture,” Kravitz says.

When it was first rolled out, that pro­gram alone saved the lives of 27 peo­ple, he says.

Atrius Health also uses nat­u­ral lan­guage pro­cess­ing to cre­ate queries and ex­tract clin­i­cal data from free-text fields, us­ing the Lin­gua­mat­ics plat­form. For ex­am­ple, it queries un­struc­tured echo re­ports to an­a­lyze heart pump func­tion to iden­tify pa­tients with a high risk of heart fail­ure.

The ini­tia­tive has en­abled the or­ga­ni­za­tion to bet­ter iden­tify gaps in care, says Craig Mon­sen, CMIO of the non­profit or­ga­ni­za­tion, which com­prises 32 physi­cian prac­tices with more than 800 doc­tors across east­ern Mas­sachusetts.

The or­ga­ni­za­tion has doc­u­mented real-world re­sults. In 2017, Atrius

Health iden­ti­fied 92 oth­er­wise un­doc­u­mented con­ges­tive heart fail­ure and chronic ob­struc­tive pul­monary dis­ease (COPD) pa­tients and gained more than $75,000 in ad­di­tional an­nual risk-ad­justed rev­enue per dis­ease area by in­ter­ven­ing.

Nurses also ben­e­fited with a sig­nif­i­cant in­crease in the ef­fi­ciency of their chart rec­on­cil­i­a­tion process, Mon­sen says. Prob­lem lists are more ac­cu­rate and com­plete, and ACO re­port­ing is sim­pli­fied.

Clos­ing gaps in care

In ad­di­tion to iden­ti­fy­ing gaps in care, pre­dic­tive an­a­lyt­ics of­fers the op­por­tu­nity to iden­tify po­ten­tially “clos­able care gaps,” Mon­sen says. From there, providers can reach out to pa­tients, such as those who are likely to miss ap­point­ments, put off tests and pro­ce­dures or fail to fol­low dis­charge plans and med­i­ca­tion reg­i­mens, and cre­ate a care plan they can en­gage with, he says.

Geisinger is also work­ing to close care gaps, es­pe­cially for pa­tients with chronic con­di­tions, such as COPD or di­a­betes, us­ing its big data plat­forms. It uses the data to cre­ate elec­tronic health record or­der sets so that the next time the pa­tient presents for care at any point in the care con­tin­uum, clin­i­cians can use those sets to take ac­tion to nar­row or close those gaps in care.

HIE data iden­ti­fies pa­tients who get care out­side of the Geisinger sys­tem to avoid cre­at­ing un­nec­es­sary or­der sets.

Pre­dic­tive an­a­lyt­ics of­fers the op­por­tu­nity to iden­tify po­ten­tially closeable caps in care.

There’s a sim­i­lar ef­fort un­der­way in In­di­ana—a pa­tient-cen­tered data hub alerts IHIE mem­bers if one of their pa­tients shows up in a med­i­cal lo­ca­tion out­side of the state. The ex­change can re­quest that clin­i­cal data and serve it up to physi­cians.

“We’re gath­er­ing data from the full land­scape of health­care,” Chris­tian says.

Pre­dict­ing read­mis­sion risk

Westch­ester Med­i­cal Cen­ter puts data an­a­lyt­ics to work to de­velop a read­mis­sion risk pre­dic­tion model.

The data an­a­lyt­ics teams at WMC and three-hos­pi­tal Bon Se­cours Char­ity Health Sys­tem, which is part of the WMC net­work, used ar­ti­fi­cial in­tel­li­gence and ma­chine learn­ing to an­a­lyze his­tor­i­cal data on pa­tient dis­charges and 30-day in­pa­tient read­mis­sions to iden­tify pa­tients who were more likely to read­mit and de­ter­mine what they had in com­mon, us­ing data an­a­lyt­ics firm Health Cat­a­lyst and the open-source ma­chine learn­ing soft­ware Health­care.ai.

The risk mod­els in­clude 24 vari­ables

that were most pre­dic­tive of read­mis­sions within 30 days, in­clud­ing age, type of di­ag­no­sis at first en­counter, and chronic con­di­tions, such as di­a­betes or COPD.

“We look for as­so­ci­a­tions and re­la­tion­ships; we run re­gres­sion mod­els to fig­ure out what fac­tors might be more likely to con­trib­ute to an out­come,” WMC’s Vi­ola says.

“So when you take this ap­proach, it can fill those crit­i­cal gaps in chronic or pre­ven­ta­tive care by fo­cus­ing your re­sources and at­ten­tion on those pa­tients who, when they’re dis­charged, may have a higher read­mis­sion risk score.

“So then you can say, ‘What is it we need to do to help that in­di­vid­ual or to en­sure that in­di­vid­u­als are not go­ing to read­mit to us, es­pe­cially if that’s not the right path­way to their care?’”

The model is 17 per­cent more ac­cu­rate than widely used read­mis­sion risk mod­els in iden­ti­fy­ing pa­tients at high risk and low risk for read­mis­sion within 30 days, in part be­cause it’s based on data from WMC’s own pa­tient pop­u­la­tion.

“The key to health sys­tems tak­ing this on and learn­ing from their own pa­tient data­bases is that we can re­ally re­flect the char­ac­ter­is­tics of our pa­tient pop­u­la­tions. We’re not re­ly­ing on these generic risk model al­go­rithms that weren’t trained on our pop­u­la­tion. So we’re more likely to cap­ture those fac­tors that put our folks at risk for pre­ventable read­mis­sion,” Vi­ola says.

“By hav­ing this risk score, you have a sense of the pa­tient that’s more ur­gent, so we can look at high-risk pa­tients and also the ris­ing-risk pa­tients,” Vi­ola says.

The cus­tom­ized risk model re­duced false pos­i­tives by 30 per­cent and also in­creased true pos­i­tives, en­sur­ing that crit­i­cal re­sources are go­ing to the right pa­tients, Vi­ola says. It also saves time—pre­vi­ously, care team mem­bers were pulling data from mul­ti­ple sources and “tri­an­gu­lat­ing” it on their own. Val­i­dated data on high- and ris­ing-risk pa­tients saved more than 1,200 hours per year.

“Af­ter we run this through and get a bet­ter sense of what vari­ables are sig­nif­i­cant for our pop­u­la­tion, we use our own an­a­lyt­ics plat­form, and we can bring in other data, such as clin­i­cal and billing data, against sub­groups of pa­tients to en­sure we’ve iden­ti­fied the right fac­tors,” Vi­ola adds. Ma­chine learn­ing en­sures the al­go­rithms con­tinue to im­prove.

More ac­cu­rate, more ac­ces­si­ble data en­ables care man­agers to fol­low up with at-risk pa­tients faster—usu­ally within seven days—to con­nect pa­tients with ap­point­ments and ser­vices to help that will help pre­vent un­nec­es­sary emer­gency de­part­ment vis­its and read­mis­sions.

“If we can au­to­mate, if we can help them so that the hours they have are with di­rect pa­tient care, that’s what we should be do­ing,” Vi­ola says.

Lead­ers who are pon­der­ing how to du­pli­cate this kind of ef­fort should “think about how they can use the tech­nol­ogy … and what in­flu­ence tech­nol­ogy can have in ad­vanc­ing the work we need to do in health sys­tems,” she adds.

Work­ing hand in hand with the clin­i­cal team is also im­por­tant. “We just help build [the mod­els], but the ac­tual val­i­da­tion and op­er­a­tional­iz­ing and in­for­ma­tion is re­ally done by those teams,” she says. “This is such a col­lab­o­ra­tive process. That’s what re­ally makes this work.”

Har­ness­ing data

Pre­dic­tive mod­els like WMC’s can also pull in data about de­ter­mi­nants of health, from de­mo­graphic and so­cioe­co­nomic in­for­ma­tion to pa­tient-gen­er­ated data—and iden­tify pa­tients who are po­ten­tially sus­cep­ti­ble to any­thing from heart fail­ure to missed ap­point­ments.

“There are a lot of pop­u­la­tions at greater risk, in part be­cause of ac­cess is­sues,” Vi­ola says. “If we can un­der­stand more about those pa­tients, we can do more. We want to make sure pa­tients get the care they need.”

Geisinger, mean­while, is us­ing its con­sid­er­able data an­a­lyt­ics power to tackle the opi­oid cri­sis in its re­gion, re­duc­ing opi­oid pre­scrib­ing rates by about 65 per­cent.

Af­ter an anal­y­sis of 942 Geisinger pa­tients who over­dosed on opi­oids found a steep in­crease in the use of acute care—es­pe­cially ex­pen­sive ED ser­vices—be­fore an over­dose, the or­ga­ni­za­tion used data from its EHR to iden­tify providers with high opi­oid pre­scrib­ing rates.

“In 2015 and 2016, we were pretty high in pre­scrib­ing opi­oids, just like other health sys­tems have been,” Kravitz con­tends. Is­su­ing 30-day opi­oid pre­scrip­tions was “hard­wired” when it comes to sur­gi­cal pa­tients—in other words, pretty much au­to­matic. “There’s your key for ad­dic­tion,” he says. “That’s scary.”

Now, a dash­board shows the data— which is vis­i­ble to all providers in the sys­tem—in­clud­ing each physi­cian’s panel size, the num­ber and per­cent­age of pa­tients who are pre­scribed opi­oids, open or­ders the pa­tient hasn’t filled and the daily mor­phine equiv­a­lent dose, for ex­am­ple. While an on­colo-

More ac­cu­rate, more ac­ces­si­ble data en­ables care man­agers to fol­low up with at-risk pa­tients faster.

gist’s panel might in­clude about 7 per­cent of pa­tients on opi­oids, the goals for oth­ers, in­clud­ing pri­mary care physi­cians, are much lower—from 1 per­cent to 2 per­cent. The sys­tem flags out­liers.

From there, the or­ga­ni­za­tion in­ten­si­fied its fo­cus on nono­pi­oid pain treat­ments for pa­tients, launch­ing aware­ness pro­grams for physi­cians and pa­tients alike. For ex­am­ple, sur­geons or sur­gi­cal staff now ex­plain the risks and al­ter­na­tive pain treat­ments to pa­tients and their fam­i­lies be­fore surgery. If a pa­tient in­sists, they get a three-day pre­scrip­tion. An en­ter­prise e-pre­scrib­ing sys­tem helped to re­duce forged or stolen pre­scrip­tions and saved a sig­nif­i­cant amount of time, money and ef­fort, he adds.

“I’m not say­ing we will never pre­scribe opi­oids again—we don’t re­strict pre­scrip­tions for stage 4 can­cer pa­tients, for ex­am­ple. But we do want to re­duce it,” Kravitz says.

Pre­dic­tive an­a­lyt­ics’ fu­ture

Geisinger and other health­care or­ga­ni­za­tion see the fu­ture of pre­dic­tive an­a­lyt­ics in ge­nomic data. It could, for ex­am­ple, help iden­tify pa­tients who have a propen­sity for opi­oid ad­dic­tion, Kravitz says.

“The in­dus­try is try­ing to iden­tify those pa­tients who have a pro­cliv­ity for be­com­ing ad­dicted be­fore they get ad­dicted,” adds IHIE’s Chris­tian.

In Ne­vada, a large pop­u­la­tion health study aims to com­bine clin­i­cal, so­cioe­co­nomic, en­vi­ron­men­tal and ge­netic screen­ing data, us­ing an­a­lyt­ics to iden­tify cor­re­la­tions that im­pact health, from opi­oid ad­dic­tion to car­dio­vas­cu­lar dis­ease.

The Healthy Ne­vada pro­gram is run by the Renown In­sti­tute for Health In­no­va­tion, a col­lab­o­ra­tion be­tween the in­te­grated Renown Health net­work based in Reno, Nev., and the non­profit en­vi­ron­men­tal arm of the Ne­vada Sys­tem of Higher Ed­u­ca­tion, the Desert Re­search In­sti­tute, which has cam­puses in Reno and Las Ve­gas.

The or­ga­ni­za­tion is gath­er­ing ge­nomic data on tens of thou­sands of state res­i­dents us­ing DNA se­quenc­ing com­pany He­lix. Par­tic­i­pants have ac­cess to the kits and their re­sults for free, so they can take ac­tion on their own by fol­low­ing up with their physi­cians.

How­ever, the pro­gram will also com­pile the dei­den­ti­fied data in an ef­fort to pre­dict who is at risk for con­di­tions and to de­velop early in­ter­ven­tion pro­to­cols. The aim is to ex­pand pre­ci­sion medicine in the state—and be­yond.

The pi­lot phase used geno­typ­ing; phase two is us­ing ex­ome se­quenc­ing, which pro­vides 100 times the amount of in­for­ma­tion pro­vided by ex­ist­ing geno­typ­ing or ar­ray tech­nolo­gies, ac­cord­ing to He­lix.

The idea is to stretch pre­dic­tive an­a­lyt­ics be­yond the hos­pi­tal or physi­cian prac­tice set­ting, says Renown Health CEO Tony Slonim, MD, and into the com­mu­nity to help pa­tients live health­ier lives us­ing data.

“Health is a state of well-be­ing— men­tally, phys­i­cally and spir­i­tu­ally,” he con­tends. “So how do you com­bine [en­vi­ron­men­tal and other] data el­e­ments in the con­text of health and well-be­ing? We are now giv­ing peo­ple the op­por­tu­nity to look ahead and not look back. And that is what drives me on this pro­ject.”

An­a­lyt­ics in ac­tion

Smaller or­ga­ni­za­tions—es­pe­cially those that don’t have the cap­i­tal to spend on pre­dic­tive an­a­lyt­ics tools, staff, soft­ware and plat­forms—are still try­ing to get their heads around the busi­ness model and what ser­vices they need to keep their pop­u­la­tions healthy, Chris­tian says. But they can start by manag­ing their own em­ploy­ees and then rolling out their find­ings to lo­cal em­ploy­ers.

If an or­ga­ni­za­tion doesn’t have in-house an­a­lyt­ics re­sources or the money to build a team, a con­tract with a data an­a­lyt­ics com­pany could help, Kravitz adds. Other op­tions in­clude open-source or crowd-sourced plat­forms, such as Health­care.ai or Hadoop.

“As health sys­tems get there, these pro­cesses and this time to de­velop these mod­els...will move more quickly.”

“As health sys­tems get there, these pro­cesses and this time to de­velop these mod­els … will move more quickly,” Vi­ola says. “At first you have to just hang in there, have pa­tience and work closely with your teams.”

Pre­dic­tive an­a­lyt­ics ef­forts are of­ten as­pi­ra­tional, in­clud­ing Geisinger’s early ex­plo­rations into ge­nomics and Healthy Ne­vada’s goals to spread its se­quenc­ing and re­search na­tion­wide. That’s good news for those or­ga­ni­za­tions that don’t have the re­sources or ex­per­tise to take the first step. That’s be­cause the ad­vanced play­ers also as­pire to help­ing their smaller coun­ter­parts.

Scaled na­tion­ally, the Healthy Ne­vada pro­gram would be size-ag­nos­tic. “We’ll take on any­body,” Slonim says.

“Hope­fully we’re cre­at­ing break­throughs that oth­ers can lever­age and uti­lize,” Kravitz says of Geisinger’s ef­forts. “Any­thing we can do to help in that area, we’re just happy to help oth­ers. We’re out for the pub­lic good.” ☐

Newspapers in English

Newspapers from USA

© PressReader. All rights reserved.