Psychiatric diseases are very heterogeneous both in clinical manifestation and etiology. With the recent rise of using machine learning techniques to attempt to diagnose and prognose these disorders, ...the issue of heterogeneity becomes increasingly important. With the growing interest in personalized medicine, it becomes even more important to not only classify someone as a patient with a certain disorder, its treatment needs a more precise definition of the underlying neurobiology, since different biological origins of the same disease may require (very) different treatments.
We review the possible contributions that machine learning techniques could make to explore the heterogeneous nature of psychiatric disorders with a focus on schizophrenia. First we will review how heterogeneity shows up and how machine learning, or multivariate pattern recognition methods in general, can be used to discover it. Secondly, we will discuss the possible uses of these techniques to attack heterogeneity, leading to improved predictions and understanding of the neurobiological background of the disorder.
In a recent review, it was suggested that much larger cohorts are needed to prove the diagnostic value of neuroimaging biomarkers in psychiatry. While within a sample, an increase of diagnostic ...accuracy of schizophrenia (SZ) with number of subjects (N) has been shown, the relationship between N and accuracy is completely different between studies. Using data from a recent meta-analysis of machine learning (ML) in imaging SZ, we found that while low-N studies can reach 90% and higher accuracy, above N/2 = 50 the maximum accuracy achieved steadily drops to below 70% for N/2 > 150. We investigate the role N plays in the wide variability in accuracy results in SZ studies (63-97%). We hypothesize that the underlying cause of the decrease in accuracy with increasing N is sample heterogeneity. While smaller studies more easily include a homogeneous group of subjects (strict inclusion criteria are easily met; subjects live close to study site), larger studies inevitably need to relax the criteria/recruit from large geographic areas. A SZ prediction model based on a heterogeneous group of patients with presumably a heterogeneous pattern of structural or functional brain changes will not be able to capture the whole variety of changes, thus being limited to patterns shared by most patients. In addition to heterogeneity (sample size), we investigate other factors influencing accuracy and introduce a ML effect size. We derive a simple model of how the different factors, such as sample heterogeneity and study setup determine this ML effect size, and explain the variation in prediction accuracies found from the literature, both in cross-validation and independent sample testing. From this, we argue that smaller-N studies may reach high prediction accuracy at the cost of lower generalizability to other samples. Higher-N studies, on the other hand, will have more generalization power, but at the cost of lower accuracy. In conclusion, when comparing results from different ML studies, the sample sizes should be taken into account. To assess the generalizability of the models, validation (by direct application) of the prediction models should be tested in independent samples. The prediction of more complex measures such as outcome, which are expected to have an underlying pattern of more subtle brain abnormalities (lower effect size), will require large samples.
Despite the multitude of longitudinal neuroimaging studies that have been published, a basic question on the progressive brain loss in schizophrenia remains unaddressed: Does it reflect accelerated ...aging of the brain, or is it caused by a fundamentally different process? The authors used support vector regression, a supervised machine learning technique, to address this question.
In a longitudinal sample of 341 schizophrenia patients and 386 healthy subjects with one or more structural MRI scans (1,197 in total), machine learning algorithms were used to build models to predict the age of the brain and the presence of schizophrenia ("schizophrenia score"), based on the gray matter density maps. Age at baseline ranged from 16 to 67 years, and follow-up scans were acquired between 1 and 13 years after the baseline scan. Differences between brain age and chronological age ("brain age gap") and between schizophrenia score and healthy reference score ("schizophrenia gap") were calculated. Accelerated brain aging was calculated from changes in brain age gap between two consecutive measurements. The age prediction model was validated in an independent sample.
In schizophrenia patients, brain age was significantly greater than chronological age at baseline (+3.36 years) and progressively increased during follow-up (+1.24 years in addition to the baseline gap). The acceleration of brain aging was not constant: it decreased from 2.5 years/year just after illness onset to about the normal rate (1 year/year) approximately 5 years after illness onset. The schizophrenia gap also increased during follow-up, but more pronounced variability in brain abnormalities at follow-up rendered this increase nonsignificant.
The progressive brain loss in schizophrenia appears to reflect two different processes: one relatively homogeneous, reflecting accelerated aging of the brain and related to various measures of outcome, and a more variable one, possibly reflecting individual variation and medication use. Differentiating between these two processes may not only elucidate the various factors influencing brain loss in schizophrenia, but also assist in individualizing treatment.
Neurobiological heterogeneity in schizophrenia is poorly understood and confounds current analyses. We investigated neuroanatomical subtypes in a multi-institutional multi-ethnic cohort, using novel ...semi-supervised machine learning methods designed to discover patterns associated with disease rather than normal anatomical variation. Structural MRI and clinical measures in established schizophrenia (n = 307) and healthy controls (n = 364) were analysed across three sites of PHENOM (Psychosis Heterogeneity Evaluated via Dimensional Neuroimaging) consortium. Regional volumetric measures of grey matter, white matter, and CSF were used to identify distinct and reproducible neuroanatomical subtypes of schizophrenia. Two distinct neuroanatomical subtypes were found. Subtype 1 showed widespread lower grey matter volumes, most prominent in thalamus, nucleus accumbens, medial temporal, medial prefrontal/frontal and insular cortices. Subtype 2 showed increased volume in the basal ganglia and internal capsule, and otherwise normal brain volumes. Grey matter volume correlated negatively with illness duration in Subtype 1 (r = -0.201, P = 0.016) but not in Subtype 2 (r = -0.045, P = 0.652), potentially indicating different underlying neuropathological processes. The subtypes did not differ in age (t = -1.603, df = 305, P = 0.109), sex (chi-square = 0.013, df = 1, P = 0.910), illness duration (t = -0.167, df = 277, P = 0.868), antipsychotic dose (t = -0.439, df = 210, P = 0.521), age of illness onset (t = -1.355, df = 277, P = 0.177), positive symptoms (t = 0.249, df = 289, P = 0.803), negative symptoms (t = 0.151, df = 289, P = 0.879), or antipsychotic type (chi-square = 6.670, df = 3, P = 0.083). Subtype 1 had lower educational attainment than Subtype 2 (chi-square = 6.389, df = 2, P = 0.041). In conclusion, we discovered two distinct and highly reproducible neuroanatomical subtypes. Subtype 1 displayed widespread volume reduction correlating with illness duration, and worse premorbid functioning. Subtype 2 had normal and stable anatomy, except for larger basal ganglia and internal capsule, not explained by antipsychotic dose. These subtypes challenge the notion that brain volume loss is a general feature of schizophrenia and suggest differential aetiologies. They can facilitate strategies for clinical trial enrichment and stratification, and precision diagnostics.
Schizophrenia spectrum disorders (SSD) have heterogeneous outcomes. If we could predict individual outcome and identify predictors of outcome, we could personalize and optimize treatment and care. ...Recent research showed that recovery rates tend to stabilize early in the course of disease. Short- to medium- term treatment goals are most relevant for clinical practice.
We performed a systematic review and meta-analysis to identify predictors of outcome ≤1 year in prospective studies of patients with SSD. For our meta-analysis risk of bias was assessed with the QUIPS tool.
178 studies were included for analysis. Our systematic review and meta-analysis showed that the chance of symptomatic remission was lower in males, and in patients with longer duration of untreated psychosis, more symptoms, worse global functioning, more previous hospital admissions and worse treatment adherence. The chance of readmission was higher for patients with more previous admissions. The chance of functional improvement was lower in patients with worse functioning at baseline. For other proposed predictors of outcome, like age at onset and depressive symptoms, limited to no evidence was found.
This study illuminates predictors of outcome of SSD. Level of functioning at baseline was the best predictor of all investigated outcomes. Furthermore, we found no evidence for many predictors proposed in original research. Possible reasons for this include the lack of prospective research, between-study heterogeneity and incomplete reporting. We therefore recommend open access to datasets and analysis scripts, enabling other researchers to reanalyze and pool the data.
•Different machine learning algorithms have been used to compute brain-predicted-age-difference (brainPAD) in schizophrenia.•Inter-study BrainPAD variability is typically attributed to ...schizophrenia-related heterogeneity rather than the algorithms used.•Comparison of 6 different algorithms in 4 independent samples yielded consistent differences in their performance.•BrainPAD estimates in the same sample of patients with schizophrenia differed by up to 10 years depending on the algorithm used.
Brain-predicted age difference (brainPAD) has been used in schizophrenia to assess individual-level deviation in the biological age of the patients’ brain (i.e., brain-age) from normative reference brain structural datasets. There is marked inter-study variation in brainPAD in schizophrenia which is commonly attributed to sample heterogeneity. However, the potential contribution of the different machine learning algorithms used for brain-age estimation has not been systematically evaluated. Here, we aimed to assess variation in brain-age estimated by six commonly used algorithms ordinary least squares regression, ridge regression, least absolute shrinkage and selection operator regression, elastic-net regression, linear support vector regression, and relevance vector regression when applied to the same brain structural features from the same sample. To assess reproducibility we used data from two publically available samples of healthy individuals (n = 1092 and n = 492) and two further samples, from the Icahn School of Medicine at Mount Sinai (ISMMS) and the Center of Biomedical Research Excellence (COBRE), comprising both patients with schizophrenia (n = 90 and n = 76) and healthy individuals (n = 200 and n = 87). Performance similarity across algorithms was compared within each sample using correlation analyses and hierarchical clustering. Across all samples ordinary least squares regression, the only algorithm without a penalty term, performed markedly worse. All other algorithms showed comparable performance but they still yielded variable brain-age estimates despite being applied to the same data. Although brainPAD was consistently higher in patients with schizophrenia, it varied by algorithm from 3.8 to 5.2 years in the ISMMS sample and from to 4.5 to 11.7 years in the COBRE sample. Algorithm choice introduces variations in brain-age and may confound inter-study comparisons when assessing brainPAD in schizophrenia.
Changes in cortical thickness over time have been related to intelligence, but whether changes in cortical surface area are related to general cognitive functioning is unknown. We therefore examined ...the relationship between intelligence quotient (IQ) and changes in cortical thickness and surface over time in 504 healthy subjects. At 10 years of age, more intelligent children have a slightly thinner cortex than children with a lower IQ. This relationship becomes more pronounced with increasing age: with higher IQ, a faster thinning of the cortex is found over time. In the more intelligent young adults, this relationship reverses so that by the age of 42 a thicker cortex is associated with higher intelligence. In contrast, cortical surface is larger in more intelligent children at the age of 10. The cortical surface is still expanding, reaching its maximum area during adolescence. With higher IQ, cortical expansion is completed at a younger age; and once completed, surface area decreases at a higher rate. These findings suggest that intelligence may be more related to the magnitude and timing of changes in brain structure during development than to brain structure per se, and that the cortex is never completed but shows continuing intelligence-dependent development.
Psychiatric prognosis is a difficult problem. Making a prognosis requires looking far into the future, as opposed to making a diagnosis, which is concerned with the current state. During the ...follow-up period, many factors will influence the course of the disease. Combined with the usually scarcer longitudinal data and the variability in the definition of outcomes/transition, this makes prognostic predictions a challenging endeavor. Employing neuroimaging data in this endeavor introduces the additional hurdle of high dimensionality. Machine learning techniques are especially suited to tackle this challenging problem. This review starts with a brief introduction to machine learning in the context of its application to clinical neuroimaging data. We highlight a few issues that are especially relevant for prediction of outcome and transition using neuroimaging. We then review the literature that discusses the application of machine learning for this purpose. Critical examination of the studies and their results with respect to the relevant issues revealed the following: 1) there is growing evidence for the prognostic capability of machine learning–based models using neuroimaging; and 2) reported accuracies may be too optimistic owing to small sample sizes and the lack of independent test samples. Finally, we discuss options to improve the reliability of (prognostic) prediction models. These include new methodologies and multimodal modeling. Paramount, however, is our conclusion that future work will need to provide properly (cross-)validated accuracy estimates of models trained on sufficiently large datasets. Nevertheless, with the technological advances enabling acquisition of large databases of patients and healthy subjects, machine learning represents a powerful tool in the search for psychiatric biomarkers.
White matter microstructure and volume show synchronous developmental patterns in children. White matter volume increases considerably during development. Fractional anisotropy, a measure for white ...matter microstructural directionality, also increases with age. Development of white matter volume and development of white matter microstructure seem to go hand in hand. The extent to which the same or different genetic and/or environmental factors drive these two aspects of white matter maturation is currently unknown. We mapped changes in white matter volume, surface area and diffusion parameters in mono- and dizygotic twins who were scanned at age 9 (203 individuals) and again at age 12 (126 individuals). Over the three-year interval, white matter volume (+6.0%) and surface area (+1.7%) increased, fiber bundles expanded (most pronounced in the left arcuate fasciculus and splenium), and fractional anisotropy increased (+3.0%). Genes influenced white matter volume (heritability ~85%), surface area (~85%), and fractional anisotropy (locally 7% to 50%) at both ages. Finally, volumetric white matter growth was negatively correlated with fractional anisotropy increase (r = -0.62) and this relationship was driven by environmental factors. In children who showed the most pronounced white matter growth, fractional anisotropy increased the least and vice-versa. Thus, white matter development in childhood may reflect a process of both expansion and fiber optimization.