Machine learning study finds standardized brain scan biomarker to detect depression with 66% accuracy

by

New Study Brings Bio­mark­ers For Depres­sion Clos­er To The Clin­ic (Forbes):

Sci­en­tists have been study­ing bio­log­i­cal signs of depres­sion in the brain, look­ing for mark­ers that could be used to iden­ti­fy the dis­or­der. A team of sci­en­tists recent­ly devel­oped a tech­nique using machine learn­ing that can iden­ti­fy whether a giv­en patien­t’s brain scan shows one of depres­sion’s neur­al sig­na­tures…

Clin­i­cians cur­rent­ly diag­nose major depres­sive dis­or­der, com­mon­ly called depres­sion, based on a patien­t’s report­ed symp­toms, like changes in mood, activ­i­ty enjoy­ment, appetite or sleep­ing, for exam­ple. But peo­ple with depres­sion can show a wide range of dif­fer­ent symp­toms, which means that depres­sion looks dif­fer­ent in each per­son … Group­ing peo­ple with wide-rang­ing symp­toms under one cat­e­go­ry also masks the fact that dif­fer­ent process­es in the brain might under­lie depres­sion in dif­fer­ent indi­vid­u­als. For these rea­sons, neu­ro­sci­en­tists have been look­ing for neur­al sig­na­tures or bio­mark­ers of depression…

Using the brain net­work bio­mark­er, their algo­rithm could cor­rect­ly iden­ti­fy which par­tic­i­pants had depres­sion 66% of the time.

While 66% accu­ra­cy may not sound high, it is an improve­ment on cur­rent accu­ra­cy lev­els of diag­no­sis by human clin­i­cians, par­tic­u­lar­ly gen­er­al physi­cians who aren’t trained in psychiatry.

The Study:

Gen­er­al­iz­able brain net­work mark­ers of major depres­sive dis­or­der across mul­ti­ple imag­ing sites (PLOS Biology):

  • Abstract: Many stud­ies have high­light­ed the dif­fi­cul­ty inher­ent to the clin­i­cal appli­ca­tion of fun­da­men­tal neu­ro­science knowl­edge based on machine learn­ing tech­niques. It is dif­fi­cult to gen­er­al­ize machine learn­ing brain mark­ers to the data acquired from inde­pen­dent imag­ing sites, main­ly due to large site dif­fer­ences in func­tion­al mag­net­ic res­o­nance imag­ing. We address the dif­fi­cul­ty of find­ing a gen­er­al­iz­able mark­er of major depres­sive dis­or­der (MDD) that would dis­tin­guish patients from healthy con­trols based on rest­ing-state func­tion­al con­nec­tiv­i­ty pat­terns. For the dis­cov­ery dataset with 713 par­tic­i­pants from 4 imag­ing sites, we removed site dif­fer­ences using our recent­ly devel­oped har­mo­niza­tion method and devel­oped a machine learn­ing MDD clas­si­fi­er. The clas­si­fi­er achieved an approx­i­mate­ly 70% gen­er­al­iza­tion accu­ra­cy for an inde­pen­dent val­i­da­tion dataset with 521 par­tic­i­pants from 5 dif­fer­ent imag­ing sites. The suc­cess­ful gen­er­al­iza­tion to a per­fect­ly inde­pen­dent dataset acquired from mul­ti­ple imag­ing sites is nov­el and ensures sci­en­tif­ic repro­ducibil­i­ty and clin­i­cal applicability.

The Study in Context: