The promise of big data analysis – precision psychiatry and therapeutic targeting

Psychiatric diagnosis - according to DSM-5 - is largely based on psychiatric symptoms, with a great deal of symptom overlap between different disorders. The brain biology responsible for any one individual’s psychiatric illness has remained elusive. Big data analysis using powerful computing technologies is, however, now revealing defects in the brain’s neuronal networks and circuits in patients with psychiatric disorders. Four pioneering studies combining temporally dense multimodal phenotyping in large samples with sophisticated computational approaches were presented in an exciting session to an enthusiastic audience by experts leading the field at APA 2018.

One size doesn’t fit all

Considerable heterogeneity and overlap is seen in psychiatric diagnosis, said Alik Sunil Widge, Assistant Professor of Psychiatry, Harvard and Massachusetts General Hospital, MA. For example, a diagnosis of major depressive disorder (MDD) can be made using 256 different combinations of symptoms.

A further limitation in psychiatric diagnosis and monitoring has been the dependence on self-reporting by patients and short consultations with the psychiatrist, he added.

To improve accuracy of diagnosis and management, it is necessary to know how patients are functioning in all of their daily activities.

Powerful computing technologies are changing the future of psychiatry

We need better tools to diagnose and monitor psychiatric illnesses, Dr Widge explained; and the solution to improving diagnosis is to apply unstructured approaches without assumptions and see how the data for different mental disorders cluster together.

Watch what people do, not what they say

Passive monitoring, using wearable technologies, can circumvent the current dependence on patient self-reporting, he added – watch what people do, not what they say, for example:

  • voice prosody, sentiment and coherence
  • activity, location, and social interactions

Interpreting rich psychiatric datasets

Bing Brunton is the first Professor of Neural Engineering, University of Washington, WA, and gave an inspirational presentation on how to apply machine learning to psychiatric datasets to uncover structure within the clustering of psychiatric symptoms.

Interpretation enables formulation of hypotheses that can then be tested in clinical trials

Professor Brunton emphasized that it is easy to accumulate large quantities of multidimensional human data, such as a huge range of physical and behavioral characteristics in many forms – numbers, integers, categories. The limiting factors are asking the right questions and who can make sense of it.

It is important to focus on what goes in and what comes out and be able to interpret the results meaningfully, she said – and then to validate the results and demonstrate their reproducibility. It is not necessary for users of algorithms to understand their mathematical basis.

Identifying data-driven types coherent across symptoms, brain and behavior, to inform treatment choice

Leanne Williams, Professor of Psychiatry and Behavioral Sciences, Stanford University, CA, highlighted the current lack of opportunities to tailor and individualize psychiatric interventions due to the current “one size fits all” approach to psychiatric diagnosis.

Big data analysis promises individualization of psychiatric interventions

Can we identify data-driven types that are coherent across symptoms, brain and behavior, and relevant to informing treatment choices? she asked.

Professor Williams and her team analyzed extensive data from 420 participants – 100 with MDD, 53 with panic disorder, 47 with post-traumatic stress disorder and 220 healthy controls.1 Integration of the data and factor analysis of symptom items identified three coherent dimensions:

  • anhedonia
  • anxious arousal
  • tension

Six cluster subtypes were also identified:

  • normative mood
  • tension
  • anxious arousal
  • general anxiety
  • anhedonia
  • melancholia

A step toward disentangling symptom overlap and identifying subtypes

These cluster subtypes clustered in symptoms profiles, but did not map onto diagnoses, and differed in cognitive behavior and brain activation.1

These findings are a step toward disentangling the symptom overlap in our current diagnoses and identifying subtypes that are coherent across specific symptoms and specific brain-behavior profiles, explained Professor Williams. This is necessary to guide tailored treatment choices.

A behavioral fingerprint that can be related to therapy

Justin Taylor Baker, Center for Law, Brain and Behavior, Massachusetts, MA, described intensive longitudinal geospatial monitoring of patients with bipolar disorder using wearable technology.

The resulting, highly-detailed individual participant timelines provide valuable insights into all activities of daily life, he explained. Robust and reproducible predictors capturing how mood and cognitive fluctuations vary from individual to individual and from episode to episode in affective and psychotic illnesses can then be developed and related to therapy.

Longitudinal data revealing the emergence of psychosis

Raquel E. Gur, Professor of Psychiatry Neurology and Radiology, University of Pennsylvania, PA, described the work she has carried out with her team on the Philadelphia Neurodevelopment Cohort (PNC).

In order to better understand the prodromal nature of early psychosis, Dr Gur’s lab undertook genotyping coupled with extensive clinical and neurocognitive phenotyping in a large sample of participants; a subsample of which also underwent neuroimaging – structural MRI, diffusion tensor imaging, functional MRI, arterial spin labeling.2-3

Dimensionality reduction subdivided symptoms into four domains – anxious-misery, fear, behavioral, psychosis. These differed by sex, in relation to each other, and to brain structure and function. Comorbidity was common in the dimensions of psychopathology.2

Participants with psychotic disorder had diminished whole-brain gray matter volume and expanded white matter volume compared with participants without significant psychopathology, Professor Gur explained. They had significantly lower gray matter volume in the medial temporal lobe and frontal, temporal, and parietal cortex; and volumetric reduction in the medial temporal lobe correlated with symptom severity.3

Professor Gur also presented her recent unpublished work demonstrating the impact of the environment on early psychosis.

Big data is here to stay

The studies presented in this symposium provide just a glimpse into the multi-faceted approaches and opportunities that big data will and has already begun to play in psychiatric diagnosis and care.


  1. Grisanzio KA et al. JAMA Psychiatry 2018;75:201-209.
  2. Shanmugan S et al. Am J Psychiatry 2016; 173:517-526.
  3. Satterthwaite TD et al. JAMA Psychiatry 2016;73:515-24