Structural neuroimaging as clinical predictor: a review of machine learning applications

1. Introduction

• Structural-MRI classifiers draw on gray-matter cortical thickness or VBM density, diffusion-derived white-matter FA, and connectivity matrices.Source: §1 · p. 2
• The review examines structural-MRI classifiers for predicting clinical state and identifying disease-related brain regions.It includes diffusion-weighted imaging techniques.Source: §1 · pp. 2–3
• Structural MRI is the focus because it offers morphological findings; other modalities are out of scope, though fMRI may be cited for generalizations.Source: §1 · p. 3
• Group-level differences need not predict individual clinical outcomes (Davatzikos, 2004; Iturria-Medina, 2013; Lo et al., 2015), so purely correlational studies are excluded.Source: §1 · p. 3
• Prediction means out-of-sample testing, including cross-validation, which the review treats as more likely to generalize.Source: §1 · p. 3
• The review centers on feature-extraction methods using relatively small feature sets and excludes deep learning because of data, compute, and interpretability demands.Source: §1 · pp. 3–4

2.1. Image processing

• T1 processing extracts cortical thickness per surface vertex or gray-matter density per voxel, while diffusion imaging can yield white-matter measures including FA, diffusivity, connectivity, and network metrics.Images are typically preprocessed and registered to a common spatial domain before tissue or structure segmentation or cortical-surface extraction.Source: §2.1 · pp. 5–6
• Computed imaging results are flattened into one feature vector per subject by removing spatial coordinates and extracting numerical values, regardless of the positions of the data points.For example, 40,000 CT values produce a 40,000×1 vector.Source: §2.1 · p. 6
• Subject vectors are aggregated into an N × M matrix, optionally incorporating demographic or behavioral data, alongside a clinical-state label for each subject.Source: §2.1 · pp. 6–7

2.2. Building a predictive model

• Because imaging datasets are often too scarce to reserve a test set, cross-validation trains on N − 1 folds and tests on the remaining fold, repeating until every fold has served as test data.Performance is reported using accuracy, sensitivity, specificity, or another suitable metric.Source: §2.2 · p. 7

2.3. Model ensembling and stacking

• Ensembling combines model predictions by weighted averaging or voting, whereas stacking feeds classifier outputs into another algorithm for the final score.These approaches can improve accuracy and reduce overfitting, though ensembles may be more robust than individual models while being harder to interpret.Source: §2.3 · p. 8
• When biological interpretability matters, researchers may prefer a slightly less accurate single model; some ensembles can also provide feature importance.Source: §2.3 · p. 8

3. Practical issues

• Internal cross-validation loops should handle feature selection and hyperparameter tuning, and permutation tests can enhance evaluation of generalizability.Performing feature selection or tuning on the full sample produces an excessively optimistic classifier.Source: §3 · p. 9, Figure 2 · p. 10

3.1. Feature preprocessing

• Removing age-related effects improved AD-versus-control accuracy by approximately 2%, and improved MCI-to-AD conversion prediction by 5%.The reported age confound was that misclassified AD patients were younger than misclassified control subjects.Source: §3.1 · p. 11

3.2. Feature selection and hyperparameter tuning

• ROI-based averaging is commonly used to reduce features from thousands to a few hundreds, but averaging over regions that are too large might lose potential differences; reducing irrelevant and redundant variables reduces computational time and improves generalization.Source: §3.2 · p. 12
• Feature selection can reveal brain regions relevant to a classification problem, complementing inferential approaches; model weights can also rank variables and map their importance.Linear SVMs and random forests can provide variable weights for feature-importance reporting.Source: §3.2 · pp. 12–13
• SVM accuracy falls short of the optimal case using only relevant features; Gaussian SVMs performed better than linear SVMs in lower-dimensional problems, while nonlinear SVMs can overfit noisy patterns.Nonlinear SVMs may achieve higher validation accuracy when dimensionality is reduced.Source: §3.2 · pp. 13–14

3.2.1. Leakage in cross-validation techniques

• Using label-related information from the full sample during feature selection leaks outcome information and can make reported accuracy overly optimistic; hyperparameter tuning should likewise occur in an inner fold loop.Selecting high-variance variables without using output labels is not considered leakage.Source: §3.2.1 · pp. 15–16

3.2.2. Bias-Variance Trade-off

• MRI feature counts often exceed subject counts, so weak regularization risks overfitting and poor out-of-sample generalization, whereas excessive regularization can underfit; the optimal balance depends on sample size and task.Source: §3.2.2 · pp. 16–17

4. Machine learning applied to structural neuroimaging

• The review examines machine-learning classifiers built for specific diseases or disorders.A few studies predicting non-categorical outputs, such as autism ADOS scores, are included as exceptions.Source: §4 · p. 17
• Classification accuracy is the proportion of predictions that are correct, counting true positives and true negatives.It is defined as the sum of true positive and true negative predictions divided by the total number of predictions.Source: §4 · p. 17
• Accuracy scores are reported on a 0-to-1 scale, rather than as percentages.For example, an accuracy of 0.67 is reported instead of a percentage.Source: §4 · p. 17

4.1. Alzheimer’s disease/ mild cognitive impairment

• Studies target AD-versus-control, AD-or-control-versus-MCI, and MCI-converter-versus-non-converter classification, covering diagnosis and prognosis.The review includes studies of sporadic AD and retains the traditional MCI-c/MCI-nc terminology for consistency with the reviewed publications.Source: §4.1 · p. 18
• AD-versus-MCI classification is harder than AD-versus-control, while predicting MCI progression within 6 months to 3 years is especially challenging (Wee et al., 2012; Westman et al., 2012).The review attributes difficulty to MCI’s diagnostic gray area and potential confusion with mild AD or healthy controls.Source: §4.1 · p. 18
• 0.89 vs. 0.86 AUC followed adding DTI to T1 in Dyrba et al. (2015), but all three modalities yielded accuracy 0.79 and AUC 0.82.T1 alone yielded accuracy 0.82; the authors suggested noisy rs-fMRI data may have caused SVM overfitting, illustrating that more features need not improve validation accuracy.Source: §4.1 · pp. 20–21
• 0.811–0.964 accuracy was reported by Klöppel et al. (2008) for controls versus AD using training and test sets from different databases, supporting cross-site generalization.The review notes that mixing images from different sites can confound results, making this validation setup a robustness test.Source: §4.1 · p. 21
• AUC = 0.95, sensitivity = 0.73, and specificity = 0.94 were obtained on a test set for controls-versus-MCI classification (Desikan et al., 2009).Training used OASIS data and testing used ADNI data.Source: §4.1 · p. 21
• 0.883 accuracy for controls-versus-AD and 0.687 for MCI-versus-AD were reported when 1.5 T and 3 T images were mixed (Schmitter et al., 2015).MCI prognosis accuracies were 0.688 at 2 years and 0.698 at 3 years; using the field-strength datasets independently yielded similar results.Source: §4.1 · p. 22
• 82.9%, 86.4%, 75.4%, and 68% of MCI subjects were identified as AD patients at 12, 18, 24, and 36 months, respectively (Westman et al.).Source: §4.1 · p. 22
• A sensitivity difference of up to 0.3 arose solely from image-processing technique (Cuingnet et al., 2011), while feature selection improved accuracy by up to 5%.Moradi et al. also reported accuracy up to 0.745 (AUC = 0.766) for MCI-nc-versus-MCI-c classification using features selected in controls-versus-AD classification.Source: §4.1 · pp. 22–23
• ACC = 0.734 was obtained for MCI-nc versus MCI-c classification using patterns derived from AD and healthy controls.Overlap in important regions was also reported (Aguilar et al., 2013; Cuingnet et al., 2011; Desikan et al., 2009; Westman et al., 2012).Source: §4.1 · p. 23
• 0.63 multi-class accuracy on the unobserved CaDD Dementia test dataset was achieved by Sørensen’s LDA classifier.The classifier distinguished controls, MCI, and AD, was trained on more than 600 subjects from multiple datasets, and its performance was examined against training-set size and classifier complexity.Source: §4.1 · p. 23
• Stacking/ensembling was uncommon; examples combine imaging output with age and behavioral data, data sources with SVM kernels, or weak classifiers.Source: §4.1 · p. 23
• Ingalhalikar et al. (2014) used classifier ensembling to handle missing data, creating classifiers per subject based on the available data subset and merging their outputs.Source: §4.1 · pp. 23–24
• Across reviewed AD classification studies, hippocampus, temporal lobes, amygdala, parahippocampal gyrus, middle temporal gyrus, entorhinal cortex, and insula are key GM regions.The table pools AD, MCI, and control cohorts; the review cautions that whole-dataset feature selection may identify group-informative variables that are not informative at the individual level.Source: §4.1 · p. 24, Table 2 · p. 25
• Although fewer studies used DTI, findings link microstructural white-matter changes and impaired connectivity to cognitive failure; FA and MD changes appear early and independently of medial temporal GM changes.These early changes are described by Fletcher et al. (2014) and Lacalle-Aurioles et al. (2016).Source: §4.1 · p. 24
• Decreased FA and increased MD have been described in cognitively normal individuals in preclinical AD, but these changes have not been used in ML classification at that stage.Source: §4.1 · p. 24

4.2. Autism

• Some studies use ADOS as a continuous score; Sato et al. (2013) found CT correlations of rLeft = 0.29, rRight = 0.072, and rBoth = 0.362.Source: §4.2 · pp. 26–27
• Zhou et al. (2014) achieved 0.70 accuracy using T1 metrics and rs-fMRI network measures, but extracted features outside the cross-validation loop, causing leakage.Source: §4.2 · p. 26
• Ecker et al. (2010a) reported CT accuracy up to 0.9 for the left hemisphere, dropping to 0.6 for the right.Source: §4.2 · pp. 26–27
• Uddin et al. (2011) also reported lateralization, analyzing each region individually.Source: §4.2 · p. 27
• Ingalhalikar et al. (2014) reported 0.83 accuracy for ASD vs. controls and 0.7 for ASD/LI vs. ASD/LI+, using ensembles to compensate for missing data.Source: §4.2 · p. 27
• ASD effects were broadly distributed, while Wee et al. (2014) found subcortical GM values more accurate than cortical GM values.Source: §4.2 · p. 27

4.3. Multiple sclerosis

• MS machine-learning applications have focused more on automatic white-matter lesion segmentation than structural-difference classification, as diagnosis can be made by detecting lesions and labeling them is challenging (García-Lorenzo et al., 2013; Lladó et al., 2012).CIS classification is more challenging, and CIS patients may develop MS within a given time window.Source: §4.3 · p. 28
• 0.96 with lesion information, 0.84 with normal-appearing GM, and 0.91 with normal-appearing WM were Weygandt et al.’s highest accuracy rates (Weygandt et al., 2011).They used T1 and T2 images to segment the brain into lesions and normal-appearing GM and WM.Source: §4.3 · pp. 28–29
• 0.87 accuracy distinguished healthy controls from early- and late-onset pediatric MS, versus 0.807 for classifying the two MS groups (Weygandt et al., 2015).Source: §4.3 · p. 29
• 0.85, 0.83, and 0.77 accuracy distinguished early versus late MS, low versus high white-matter lesion load, and benign versus non-benign MS, respectively, using GM alone (Bendfeldt et al., 2012).The classifiers used T1 and T2 data as inputs for linear SVMs.Source: §4.3 · p. 29
• 0.714 accuracy at 1 year and 0.68 at 3 years showed that longer-horizon prediction of MS from CIS was harder (Wottschel et al., 2015).Optimal features differed by horizon: 1-year prediction used lesion load, presentation type, and gender; 3-year prediction used lesion measures plus age and EDSS at onset.Source: §4.3 · p. 29
• The middle frontal gyrus was most informative for Weygandt et al. (Weygandt et al., 2015), while Bendfeldt et al. found relevant cortical regions across all lobes, plus thalamus and caudate (Bendfeldt et al., 2012).Source: §4.3 · p. 29

4.4. Parkinson’s disease and related disorders

• Reviewed tasks include PD, PSP, MSA variants, and SWEDD, whose subjects show PD symptoms without PET evidence of dopaminergic deficit, as well as healthy controls and combinations.PSP and MSA share motor symptoms with PD but progress faster and respond less to treatment.Source: §4.4 · p. 30
• WM voxel features yielded high control-versus-PSP and PD-versus-PSP accuracy in Focke et al. (2011), while Cherubini et al. (2014) reported WM-only accuracy = 1.GM values yielded much lower accuracies in Focke et al.; both studies used whole-sample F-contrast values as input weights outside the cross-validation loop.Source: §4.4 · p. 31
• F-contrast weights computed on the whole sample outside cross-validation could cause leakage, and smoothed WM signal may include nearby GM nuclei.The reported WM areas were mainly in the brainstem; smoothing can place small nuclei such as the substantia nigra pars compacta within the WM mask.Source: §4.4 · pp. 31–32
• Filippone et al. (2012) used multinomial logit on 62 subjects, while Marquand et al. (2013) varied control inclusion and split MSA into MSA-P and MSA-C.The Filippone cohort comprised 14 healthy controls, 14 PDs, 16 PSPs, and 18 MSAs.Source: §4.4 · p. 32
• Including healthy controls in Marquand et al.’s multiclass environment lowered overall accuracy scores; Focke et al. attributed this to VBM-processing inconsistencies.Source: §4.4 · p. 32
• The review concludes that PD, PSP, and MSA affect different brain regions despite similar symptoms.Source: §4.4 · p. 32, Table 6 · p. 33

4.5.1. Attention deficit hyperactivity disorder

• ADHD studies reported accuracy 0.611 without statistical significance, 0.793 for 29 ADHD patients and 29 healthy controls (Lim et al., 2013), and up to 0.902 using extreme learning, with feature selection outside cross-validation.Source: §4.5.1 · pp. 33–34

4.5.3. Schizophrenia

• For schizophrenia, Zarogianni et al. (2013) reviewed studies reporting 81–91.8% accuracy for sMRI classification of patients versus healthy controls; fewer transition-prediction studies reported accuracies of 61.8–95%.Source: §4.5.3 · p. 35
• Fewer schizophrenia studies addressed outcomes: one reported 85% accuracy for identifying treatment responders using EEG, and three reported 82–84.2% for transition prediction; reviewers attributed easier case-control classification to more distinct patterns.The cited transition studies used SVMs; the responder study used kernel partial least squares regression.Source: §4.5.3 · pp. 35–36

4.5.4. Traumatic brain injury

• Two TBI studies used whole-sample mRMR feature selection before cross-validation (leakage); reported accuracies were 0.86 for multilayer perceptron with selected variables, 0.80 for Bayesian network with all variables, and 0.934 for linear SVM.Lui et al. used T1, DTI, and resting-state fMRI; Fagerholm et al. used DTI graph metrics.Source: §4.5.4 · p. 36

4.5.5. Stroke

• In stroke, DTI and resting-state fMRI classified patients versus normal controls with accuracy 0.826 and predicted motor impairment with accuracy 0.876; DWI lesion maps distinguished patients with versus without hand motor impairment with accuracy 0.738 and sensitivity 0.50.Source: §4.5.5 · p. 36

4.5.6. Miscellanea

• Lavagnino et al. (2015) reported 0.833 accuracy classifying 15 anorexia nervosa patients and 15 healthy controls with LASSO using FreeSurfer-processed T1 information; Hajek et al. (2015) reported 0.689 with linear SVM (C=1) distinguishing 45 healthy subjects from 45 high-risk offspring of BD subjects.Source: §4.5.6 · p. 37

5. Discussion

• The review describes separate binary classifiers for multiple pairwise comparisons as the common practical approach to intermediate or related clinical states.Each classifier’s performance is assessed individually.Source: §5 · p. 38
• Only a few studies used multiclass methods, but these may be more clinically informative because binary classifiers exclude other labels and may miss spectrum-disorder subtleties.The cited multiclass studies are Filippone et al. (2012) and Marquand et al. (2013).Source: §5 · p. 38
• Across diseases including AD, autism, MS, and PD, normal-control versus patient classification accuracies were higher than 0.9, while progressive-versus-stable patient accuracies were generally around 0.7, despite greater clinical interest.Source: §5 · p. 39
• Clinical trials that predict outcomes before they occur can confirm prognosis models, while externally administered challenges with hidden outcomes can test against leakage and overfitting.The review cautions that generalizations from small samples and few studies require care.Source: §5 · p. 39
• Structural MRI patterns appear especially indicative of AD, whereas schizophrenia studies’ reliance on functional modalities may signal structural MRI’s limits; AD studies also generally have larger samples.The authors caution that these disease-level generalizations draw on differing numbers of studies and sample sizes.Source: §5 · pp. 39–40
• This review focuses on structural MRI, although other imaging and clinical measures may add informative features; baseline clinical information improved converter-versus-non-converter MCI prediction.Adding modalities does not always improve accuracy, since extra features can introduce noise and confounds.Source: §5 · p. 40
• Clinical labels may imperfectly approximate subject state or contain diagnostic errors, while outcomes expressed as rates or amounts of change may be better modeled with regression than binary classification.The review itself focuses on predictions formulated as binary classification tasks.Source: §5 · p. 41
• SVMs predominate in neuroimaging prediction, and the review notes their robust behavior when variables greatly outnumber available subjects.Source: §5 · p. 42

• Other techniques have produced comparable results, and feature selection can reduce dimensionality when applied correctly to avoid leakage.Source: §5 · p. 42
• Selecting informative variables outside the cross-validation loop, including through test-set labels, can bias results through leakage.The review notes that some studies selected features using group differences in the full sample or other filtering procedures before cross-validation.Source: §5 · pp. 42–43
• Leakage is especially consequential in small samples: nearly all reviewed studies had fewer than 250 instances, though the threshold depends on classifier and feature count.Source: §5 · p. 43
• Because studies use different populations and features, meaningful cross-study accuracy comparisons are not feasible, leaving leakage-related accuracy differences unassessed.Source: §5 · p. 43
• Scanner and acquisition heterogeneity limits generalizability: classifier performance may be reliable only on data collected with scanners and parameters similar to training data.Source: §5 · p. 43
• Preprocessing tools such as SPM, FSL, and MINC address variability, while multi-site, multi-scanner datasets such as ADNI and PPMI are used to improve generalizability (Aubert-Broche et al., 2013; Jenkinson et al., 2012; Penny et al., 2011).Source: §5 · p. 43
• In the ADHD-200 Global Competition, the highest accuracy (0.625) came from age, sex, handedness, and IQ alone, without imaging.Source: §5 · p. 45