Poor per­for­mance

Measly gains from mod­els that pre­dict read­mis­sions

Modern Healthcare - - THE WEEK IN HEALTHCARE - Melanie Evans

es­earchers on the hunt for clues to pre­dict which hos­pi­tal pa­tients will re­turn for a sec­ond stay have so far made only mar­ginal progress, said au­thors of a newly pub­lished study.

Re­searchers found stud­ies of mod­els that could point to pa­tients at high risk of re­peat hos­pi­tal vis­its—us­ing vari­ables such as age, di­ag­no­sis, prior med­i­cal care—have been mostly lim­ited and not ter­ri­bly ef­fec­tive.

“The thing that was ev­i­dent from re­view­ing a cou­ple dozen of these mod­els was how com­plex read­mis­sion risk re­ally is,” said Dr. De­van Kansagara, the study’s lead author and di­rec­tor one of four ev­i­dence-based syn­the­sis pro­grams for the Vet­er­ans Af­fairs Depart­ment. Agency of­fi­cials rely on anal­y­sis from the four pro­grams to help guide VA pol­icy.

The study was pub­lished in the Oct. 19 is­sue of the Jour­nal of the Amer­i­can Med­i­cal As­so­ci­a­tion.

Kansagara and six other re­searchers re­viewed 30 stud­ies of 26 mod­els to pre­dict hos­pi­tal read­mis­sions. “One of the bot­tom lines is that this is a com­plex phe­nom­e­non and the fac­tors ex­tend well be­yond their ill­ness and co-mor­bid­ity,” he said.

For hos­pi­tals, an ef­fec­tive model may bet­ter iden­tify pa­tients for pro­grams to pre­vent avoid­able hos­pi­tal vis­its.

Fed­eral of­fi­cials have iden­ti­fied read­mis­sion rates as a mea­sure of qual­ity and have tied pay­ment penal­ties to high rates start­ing

Rin 2013. Four­teen stud­ies re­viewed by Kansagara and his col­leagues re­lied on ret­ro­spec­tive data. Of those, nine were large U.S. stud­ies that gen­er­ally per­formed poorly, in­clud­ing three CMS stud­ies for con­ges­tive heart fail­ure, acute my­ocar­dial in­farc­tion and pneu­mo­nia.

With­out more ac­cu­rate mod­els, Kansagara said, hos­pi­tals risk wast­ing re­sources by tar­get­ing the wrong pa­tients. Find­ings also raise ques­tions whether pub­lic com­par­isons for pay­ment in­cen­tives could be flawed, with un­in­tended con­se­quences, he said.

One com­par­i­son found no dif­fer­ence be­tween a pre­dic­tive model that con­sid­ered sev­eral fac­tors—age, sex, self-re­ported health, heart dis­ease or di­a­betes di­ag­no­sis, prior med­i­cal use, prior hos­pi­tal­iza­tion and help from an in­for­mal care­giver—and the best guess of doc­tors and med­i­cal res­i­dents and in­terns.

The model and doc­tors proved poor pre­dic­tors of who would land back in the hos­pi­tal. Nurses and case man­agers had even less suc­cess.

John Adams, a se­nior statis­ti­cian in RAND Health, said he was not sur­prised by the find­ings. “It’s just plain hard,” he said of de­vis­ing an ac­cu­rate model. “Pre­dic­tions are dif­fi­cult.” Adams said he also un­der­stands the in­ter­est and hopes that pre­dic­tive health­care mod­els in­spire, de­spite their mostly dis­ap­point­ing per­for­mance so far.

How­ever, Adams noted that sta­tis­ti­cal scores com­monly used to mea­sure pre­dic­tive mod­els’ ac­cu­racy should not be the only cri­te­ria for their use. Even mod­er­ately ac­cu­rate mod­els may be use­ful to tar­get pre­ven­tion and case-man­age­ment pro­grams if the cost is off­set by the gains, he said.

Re­search has ex­plored a lim­ited num­ber of po­ten­tially telling in­di­ca­tors, Kansagara said.

Most stud­ies con­sid­ered whether pa­tients had mul­ti­ple dis­eases and how of­ten, if at all, pa­tients used med­i­cal care or were hos­pi­tal­ized, the au­thors wrote. Nearly all stud­ies also used pa­tients’ age and gender as fac­tors to pre­dict fu­ture hos­pi­tal vis­its.

But other vari­ables were fre­quently ab­sent from re­search. Of­ten omit­ted were mea­sures of pa­tients’ over­all abil­ity to func­tion, the sever­ity of their ill­ness, and fac­tors such as in­come, so­cial sup­port and ac­cess to care.

Kansagara also had a prac­ti­cal mo­tive for the re­search. He and a col­league have be­gun to de­sign a pro­gram to pre­vent un­nec­es­sary hos­pi­tal stays for vul­ner­a­ble pa­tients at the Port­land Vet­er­ans Af­fairs Med­i­cal Cen­ter. With no highly ac­cu­rate model to help find high-risk pa­tients or tar­get pre­ven­tion ef­forts, the physi­cians in­stead will care­fully con­sider the char­ac­ter­is­tics of pa­tients they hope to help, he said. “There is no off-theshelf-model.”

Newspapers in English

Newspapers from USA

© PressReader. All rights reserved.