By Marilynn Larkin
NEW YORK—MRI-based predictive tools could inform therapeutic decision making, potentially preventing functional impairment in patients at high clinical risk of psychosis and those with recent-onset depression, researchers say.
"Psychiatry, in contrast to oncology, for example, currently does not have any quantitative tools to predict critical outcomes at the single-patient level," Dr. Nikolaos Koutsouleris of Ludwig-Maximilian-University in Munich, Germany, told Reuters Health.
"We were able to show that machine learning can be used to prognosticate outcomes (such as) social and vocational disability in adolescents and young adults at-risk of developing psychiatric illnesses - specifically, psychosis or relapsing depression," he said by email.
"Our prediction algorithms showed very good generalizability across seven European countries, which is particularly relevant for the clinical application of these tools," he noted.
"Furthermore," he said, "we generated a prognostic workflow to achieve robust prediction, starting with easy-to-obtain clinical data on the patients' functioning (on the job, in high school and socially) and then adding structural brain imaging data (MRI) to improve prediction in cases who could not be predicted with sufficient reliability using clinical data alone."
Dr. Koutsouleris and colleagues studied 116 individuals at clinical high risk (CHR) for psychosis (mean age, 24; 50% women) and 120 with recent-onset depression (ROD; mean age, 26.1; 54.2% women). Participants were followed for a mean of 329 days.
As reported online September 26 in JAMA Psychiatry, using clinical baseline data, machine learning predicted the 1-year social-functioning outcomes with a balanced accuracy of 76.9% in CHR patients and 66.2% in those with ROD.
Balanced accuracy in models using structural neuroimaging was 76.2% in CHR patients and 65% in those with ROD; in models that combined imaging with clinical data, it was 82.7% for CHR and 70.3% for ROD.
Lower functioning before study entry was a transdiagnostic predictor.
In the CHR group, medial prefrontal and temporo-parietooccipital gray matter volume (GMV) reductions and cerebellar and dorsolateral prefrontal GMV increments had predictive value, whereas reduced mediotemporal and increased prefrontal-perisylvian GMV had predictive value in patients with ROD.
In CHR patients but not those with ROD, poor prognoses were associated with increased risk of psychotic, depressive, and anxiety disorders at follow-up.
Overall, machine learning outperformed expert clinical prognostication.
Further, adding neuroimaging machine learning to clinical machine learning provided a 1.9-fold increase in prognostic certainty in uncertain CHR cases, and a 10.5-fold increase in prognostic certainty for those with ROD.
"Clinicians could use these tools to obtain decision support before starting treatment in a given patient," Dr. Koutsouleris said. "A patient with a poor prognosis could receive more intensive psychological / psychosocial treatments to avert functional deterioration, while patients with a good prognosis could be included in rather lenient follow-up schemes."
"My experience is that adding a quantitative basis to clinical decision making helps the affected individuals become much more receptive to the concept of prevention," he noted. "This improves treatment adherence and communication between all sides involved in the therapeutic process. It mitigates the subjectivity problem still inherent in all psychiatric treatment decisions."
"The next next step," he added, "is to test these predictive algorithms in clinical practice on a large scale to ascertain the clinical and health economic benefits of these new quantitative tools. To facilitate these studies, we are working on making our tools available through the webpage http://proniapredictors.eu/. "
Dr. Aristotle Voineskos of the University of Toronto, author of a related editorial, commented, "These findings clearly represent a large body of work and effort, and in particular support the value of collaborative multi-center research combining clinical and neuroscience measures."
"The emphasis on prediction of functional outcome was important," he told Reuters Health, "because people with lived experience, especially youth and emerging adults, tell us that functioning is most important to them, over and above more clinician-driven measures of symptoms, etc."
"One of the more interesting findings is that the MRI scan data were better than clinician assessment in predicting functional outcome in a subset of patients," he said by email. "This finding should bring us closer to arguing for clinical utility."
"Additional work that needs to happen is to determine the cost effectiveness of doing an MRI for this subset of patients," he said, "and how it might offset inefficient use of dollars related to services a person may or may not need depending on their outcome trajectory."
In addition, he noted, "people with lived experience and family members (need) to understand their own interest in participating in an MRI to help determine features of their care, as this is not routinely done currently in mental health care."
Further, he said, "We will need to train psychiatrists and likely other providers to understand when to order an MRI, how to communicate findings, and how to incorporate findings into a care pathway."
SOURCE: http://bit.ly/2y50kFx and http://bit.ly/2y5g3nY
JAMA Psychiatry 2018.
(c) Copyright Thomson Reuters 2018. Click For Restrictions - https://agency.reuters.com/en/copyright.html