Classifying schizophrenia patients and healthy individuals: Whole brain SPECT functional connectivity using support vector machine classification

Abstract

Background

Functional magnetic resonance imaging (fMRI) has been used to characterize functional brain networks in disorders ranging from depression to schizophrenia. Like fMRI, single photon emission computed tomography (SPECT) is a technique which captures information about neurally linked blood flow activity through radioactive tracers. While a few SPECT studies in schizophrenia populations have been conducted along with fMRI based studies, research on SPECT data for individual subject classification is limited. We used an independent component analysis (ICA) approach to estimate covarying SPECT networks from our prior study. Results were then fed as input to a classifier model to evaluate accuracy of individual diagnostic prediction.

Methods

213 subjects (137 schizophrenia patients and 76 healthy controls) were used for the analysis. Classification input was based on loading parameters generated from spatially constrained ICA using a set of network priors derived from fMRI. Fifty-three SPECT components were estimated guided by the NeuroMark fMRI 1.0 template. We initially focused on a support vector machine (SVM) classifier given previous favorable fMRI-SVM results. We also evaluated performance of multiple classifiers post hoc.

Results and conclusion

Surprisingly, linear SVM performed worse compared to random forest, logistic regression, voting, and multilayer perceptron. Linear SVM had a greater number of false negatives compared to the other classifier approaches. By contrast, random forest and logistic regression performed the highest, with an 88% sensitivity and 61% specificity (random forest), and 87% sensitivity and 68% specificity (logistic regression). These results demonstrate that other approaches such as random forest and logistic regressions should be further investigated for future SPECT studies, and coupled with sc-ICA, provide an improved understanding of aberrant functional networks in schizophrenia.

Glossary

AUC –

area under the receiver operating characteristic (ROC) curve

EEG –

electroencephalography

fMRI –

functional magnetic resonance imaging

ICN –

intrinsic connectivity network

ICA –

independent component analysis

sc-ICA –

spatially constrained ICA

SPECT –

single photon emission computerized tomography

SVM –

support vector machine

1. Introduction

Schizophrenia is a complex, multifaceted disorder with patients experiencing a variety of behavioral and cognitive symptoms, including hallucinations and delusions (Rootes-Murdy et al., 2022; Silverstein and Lai, 2021). More importantly, challenges continue to persist in determining an exact cause for the onset of schizophrenia. Theories related to functional dysconnectivity among brain regions in schizophrenia have been discussed as a common deficit in patients. Dysconnectivity has been linked to key neurotransmitter imbalances in schizophrenia (Stephan et al., 2009) as well as developmental abnormalities (Harikumar et al., 2023; Karlsgodt et al., 2008; Schmitt et al., 2011). Disruptions in various functional networks have been noted in schizophrenia studies, namely disruptions in the frontoparietal attention network (Roiser et al., 2013), cerebellar and sensorimotor domains (Fu et al., 2021), and default mode network (Weber et al., 2020).

Functional neuroimaging has been used to study whole brain networks linked to neural activity. Single photon emission computed tomography (SPECT) has been used to measure blood perfusion and metabolic status (Catafau, ; Santra & Kumar, 2014), as well as to understand brain activity in psychiatric disorders (Amen et al., 2011; Amen and Easton, 2021; Kalyoncu, 2021). SPECT can be used to capture brain activity linked to blood perfusion in areas of the brain via radioactive tracers (Davis et al., 2020). Research studies show translational benefits of using SPECT in various psychiatric populations, notably that SPECT can detect biochemical changes in disorders such as depression and schizophrenia (Kalyoncu, 2021; McCutcheon et al., 2021; Percie Du Sert et al., 2023). Regional cerebral blood flow (rCBF) can be measured using SPECT through a variety of tracers, which can be utilized to measure strokes, seizures, infections, and illnesses based on the blood flow and distribution of tracer throughout the body (Crişan et al., 2022). Prior SPECT studies have shown the usefulness of various tracer types identifying group differences in schizophrenia and Parkinson's disease (Harikumar et al., 2025; Premi et al., 2017), making SPECT a useful modality in identifying group differences between case/controls. Additionally, there is evidence that SPECT can be used in conjunction with clinical assessments to narrow differential diagnoses and ensure greater accuracy in diagnosing patients compared to using assessments alone (Amen and Easton, 2021; Best, 2022; Malhotra et al., 2006). These approaches become useful in understanding complex neuropsychiatric disorders such as schizophrenia.

While a handful of studies utilizing SPECT in schizophrenia populations have been conducted (Cumming et al., 2021; Dabiri et al., 2022; Fond et al., 2021; Kalyoncu, 2021; Markos et al., 2024), there have been no studies focused on individual subject prediction from brain networks derived from SPECT data. The current schizophrenia literature has begun to take advantage of machine learning approaches more broadly to accurately detect and classify brain regions (Jimenez-Mesa et al., 2023) and, more specifically, investigate networks in schizophrenia (Chen et al., 2023; Saha et al., 2024) through the use of deep learning approaches (Popov et al., 2024). Previous research has shown that a blind source separation technique known as independent component analysis (ICA) can robustly identify intrinsic connectivity networks with high accuracy (ICNs) (Duda et al., 2023; Salman et al., 2019). Several imaging studies have used ICA for classification of schizophrenia patients vs. controls (Arbabshirani et al., 2014, 2017). Additional studies have also shown that when using classification methods such as a convolutional neural network (CNN) combined with ICA, overall classification accuracy results improved as well up to over 85% (Niu et al., 2019).

In addition to ICA and classifier methods such as autoconnectivity and CNN, an important classification approach that has yielded excellent results is the use of support vector machine (SVM) classification. SVM has been used in a variety of studies (Steardo et al., 2020) to accurately map brain networks in schizophrenia. A comprehensive review by Steardo and colleagues noted that up to 85% accuracy has been achieved for diagnostic discrimination between patients/controls using SVM in schizophrenia datasets, further highlighting the utility of using SVM for characterizing network differences (Steardo et al., 2020). Additionally as detailed above, the combined use of a classifier such as SVM + ICA have proven to be especially useful in improved classification accuracy. When investigating fusion datasets comprised of genetic data and fMRI data, results revealed that joint ICA and classifier approaches with SVM showed accuracy scores of 74% with single nucleotide polymorphism (SNP)-support vector machine ensemble (SVME), 83% for ICA – single vector machine classifier (SVMC), and 87% for combined SNP-fMRI (Yang et al., 2010). Taken together, these results with SVM alone and combined ICA + SVM studies have shown the promising utility of a combined approach in improving single subject classification in schizophrenia data.

Performing ICA on SPECT data involves certain considerations and assumptions. Given that SPECT ICA is computed using a subject x component mixing matrix (unlike fMRI, which uses time courses and utilizes a time x component matrix; Esposito, 2003), there are no functional network connectivity (FNC) measures to compute. Instead, subject x component connectivity matrices are created. Furthermore, given the lack of studies using single-subject classification on SPECT data, we utilize a fully automated fMRI-guided ICA approach called NeuroMark that leverages spatially constrained independent component analysis (sc-ICA) to constrain the SPECT components to 53 brain regions of interest (Du et al., 2020a, Du et al., 2020b, Table 1a, Table 1b). sc-ICA uses spatial priors derived from replicable brain networks and estimates subject specific expressions of independent SPECT components (ICs) from the SPECT data. Next, we utilize and focus on the results from the linear SVM approach, along with 8 additional classification methods posthoc to compare results. We specifically focus on SVM given robust prior fMRI findings to evaluate the single subject predictive accuracy of the SPECT data. Prior studies have used SVM to identify functional connectivity patterns in schizophrenia including relationships between the default mode network (DMN) and frontal networks with cognitive dysfunction, as well as other relationships with schizophrenia symptoms (Jing et al., 2019; Rampisela and Rustam, 2018; Sadeghi et al., 2022; Steardo et al., 2020; Verma et al., 2023). Thus, our primary objective in this study is to use fMRI guided networks from the sc-ICA analysis to classify patients from controls in SPECT data, and see how the results fare when compared to previous fMRI sc-ICA based studies. We were also interested in visualizing which networks/regions in the SPECT data were most informative regarding the prediction between patients and controls, and discuss this in further detail.

Table 1a.

Demographic table for patients and controls across the sites listing age, gender, and race information.

Amen Clinics Site #	Age (Mean | SD)	Subject Breakdown	Gender	Categorization
1	25 | 8.48 | 2 subject	2M	N/A	1 Hispanic/Latino, 1 African/American
2	31.77 | 11.08 | 35 subjects	17F/18M	48%/51%	1 Arab, 3 Asian, 19 Caucasian, 5 Hispanic/Latino, 1 Multiethnic, 1 Other, 5 Unknown
3	38.5 | 12.58 | 12 subjects	2F/10M	16.6%/83.3%	11 Caucasian, 1 Unknown
4	33.14 | 13.67 | 27 subjects	3F/24M	11.1%/88.88%	2 African/American, 23 Caucasian, 1 Hispanic/Latino, 1 Unknown
5	37.66 | 16.46 | 12 subjects	6F/6M	50%/50%	2 Asian, 8 Caucasian, 1 Native American, 1 Other
6	25.77 | 2.90 | 9 subjects	6F/3M	66.6%/33.3%	6 Caucasian, 1 Hispanic/Latino, 1 Other, 1 Unknown
7	34.10 | 15.32 | 19 subjects	4F/15M	21.05%/78.94%	1 African, 1 Asian, 15 Caucasian, 2 Unknown
8	32.54 | 11.08 | 11 subjects	6F/5M	54.54%/45.45%	2 Asian, 6 Caucasian, 1 Hispanic/Latino, 1 Other, 1 Unknown
9	30.16 | 10.22 | 6 subjects	2F/4M	33.33%/66.66%	4 Caucasian, 1 Multiethnic, 1 Unknown
10	69 | N/A (just 1 subject)	1F	N/A	1 Caucasian
11	33 | N/A (just 1 subject)	1F	N/A	1 Caucasian
12	43 | 22.62 | 2 subjects	1F/1M	50%/50%	2 Caucasian
	Total: 137 patients
2	36 | N/A | 1 subject	1M	N/A	African American
2	48.71 | 6.92 | 7 subjects	4M/3F	57.14%/42.85%	Arab/Middle Eastern
2	29.28 | 6.92 | 7 subjects	3M/4F	42.85%/57.14%	Asian
2	43.74 | 17.94 | 55 subjects	32F/23M	58.18%/41.81%	Caucasian
2	35.33 | 9.266 | 6 subjects	3F/3M	50%/50%	Hispanic/Latino
	Total: 76 controls

Table 1b.

Dosage information and calibration according to adult participant's weight; dosage amounts are in millicuries (mCi).

Weight Range	Dosage Information of TC-99
<150-175 lbs	20 mCi
175-200 lbs	22.5 mCi
>200 lbs	25.0 mCi

2. Method

213 subjects (137 schizophrenia patients and 76 healthy controls) were used for the analysis. We provide a demographic of subjects in both healthy controls and patients (Table 1a). Race classification was done based on the Center for Disease Control's guidelines for race reporting (https://www.cdc.gov/nchs/hus/sources-definitions/race.html).

Data was preprocessed before classification, and the labels were converted from categorical to numerical form. SPECT images were acquired from Amen Clinics (https://www.amenclinics.com/), with a summary of preprocessing methods detailed here from (Harikumar et al., 2025). Healthy participants were recruited and approved by Western IRB (WIRB # 20021714), with all participants consented (Integ Review Board #004-Amen Clinics Inc). Participants were not included in the original study if they, or their first-degree relatives, had thoughts of suicide or suicidal behavior, unstable medical conditions, epilepsy, a history of mental illness, head injury or other neurological disorders, dementia, intellectual disability, or substance abuse/dependence disorders. All twelve sites used two camera types (i.e. double vs. triple headed gamma cameras) and scans were performed with the same data acquisition parameters in line with Amen Clinics regulatory procedures.

Adult participants were scanned with the dosage amounts indicated below, which were calibrated by weight (Table 1b).

Participants were scanned with the same procedures listed below here across the twelve sites. Scanning was completed prior to any diagnoses or treatment. For each procedure, an age- and weight-appropriate dose of 99mTc–hexamethylpropyleneamine oxime (HMPAO) was administered intravenously at rest. Patients were injected while they sat in a dimly lit room with their eyes open. Each patient participated in two SPECT brain scans, acquired during rest and while performing a Conners Continuous Performance Test (Conners Continuous Performance Test, CCPT-II, Multi-Health Systems, Toronto, Ontario (Conners, 2000)) across twelve clinical imaging sites. Scans were performed between 2004 and 2022. Scans were performed using a Picker (Philips) Prism XP 3000 double-headed gamma camera (Picker Int. Inc., Ohio Nuclear Medicine Division, Bedford Hills, OH, USA) along with the PRISM 3000 triple-headed gamma camera (Picker Int. Inc., Ohio Nuclear Medicine Division, Bedford Hills, OH, USA) with low energy high resolution fan beam collimators. Patients were scanned for approximately 30 min after injection. 120 images per scan were generated with each image separated by three degrees, spanning 360°. A low pass filter was applied with a high cutoff and Chang attenuation correction performed (L.-T. Chang, 1978)^,(W. Chang et al., 1984). The resulting reconstructed image matrices were 128 × 128 × 78 with voxel sizes of 2.5 mm³.

For voxel-based analyses, images were aligned to the Montreal Neurological Institute (MNI) space with the advanced normalization tools (ANTs version 2.2.0 (Avants et al., 2011), RRID:SCR_004757) using a SPECT template, resulting in an image matrix size of 79 × 96 × 68 with isotropic voxel sizes of 2.0 mm³. SPECT images were scaled to the within-scan maximum voxel and noise outside of the brain was removed using a 50% of maximum threshold, prior to registration. Post thresholding, images were aligned to the MNI space and the transformation applied to the un-thresholded images. Then, a brain mask derived from the MNI 152 (V. S. Fonov et al., 2009; V. Fonov et al., 2011) template was used to remove noise outside the brain from the un-thresholded images for use in the statistical models. Registered SPECT scans were visually checked for the absence of severe functional abnormalities or artifacts and registration to the MNI space.

2.1. ICA processing

SPECT data post preprocessing were run through a spatially constrained ICA using the NeuroMark ICA template through the Group ICA of fMRI Toolbox (GIFT; http://trendscenter.org/software/gift; RRID:SCR_001953) run with MATLAB R2020b (Version 9.9.0.2037887). SPECT scans from 213 subjects total (healthy controls and patients combined) were analyzed via the MOO-ICAR algorithm using the NeuroMark fMRI 1.0 template (Table 2) which included 53 ICs that replicated across two large scale human fMRI datasets (see Du et al., 2020a, Du et al., 2020b for more about the NeuroMark template). 53 SPECT ICs as well as subject specific loading parameters were generated; consequently, these 53 ICs were spread across 7 intrinsic connectivity network (ICN) domains including the subcortical (SC), auditory (AUD), visual (VIS), sensorimotor (SM), cognitive control (CC), default mode network (DM), and cerebellar (CB) component regions.

Table 2.

Adapted from Du et al., 2020a, Du et al., 2020b) with the breakdown of coordinates, and 53 NeuroMark ICA brain regions used for the spatially constrained ICA analysis. The 7 ICNs of interest are defined here.

Independent Components as Regions of Interest (Peak Coordinates and Spatial Maps) Peak Coordinates of Components of Interest
Selected Components as Regions of Interest	X	Y	Z
Subcortical network (SC)
Caudate (1)	6.5	10.5	5.5
Subthalamus/hypothalamus (2)	−2.5	−13.5	−1.5
Putamen (3)	−26.5	1.5	−0.5
Caudate (4)	21.5	10.5	−3.5
Thalamus (5)	−12.5	−18.5	11.5
Auditory network (AUD)
Superior temporal gyrus ([STG], 6)	62.5	−22.5	7.5
Middle temporal gyrus ([MTG], 7)	−42.5	−6.5	10.5
Sensorimotor network (SM)
Postcentral gyrus ([PoCG], 8)	56.5	−4.5	28.5
Left postcentral gyrus ([L PoCG], 9)	−38.5	−22.5	56.5
Paracentral lobule ([ParaCL], 10)	0.5	−22.5	65.5
Right postcentral gyrus ([R PoCG], 11)	38.5	−19.5	55.5
Superior parietal lobule ([SPL], 12)	−18.5	−43.5	65.5
Paracentral lobule ([ParaCL], 13)	−18.5	−9.5	56.5
Precentral gyrus ([PreCG], 14)	−42.5	−7.5	46.5
Superior parietal lobule ([SPL], 15)	20.5	−63.5	58.5
Postcentral gyrus ([PoCG], 16)	−47.5	−27.5	43.5
Visual network (VIS)
Calcarine gyrus ([CalcarineG], 17)	−12.5	−66.5	8.5
Middle occipital gyrus ([MOG], 18)	−23.5	−93.5	−0.5
Middle temporal gyrus ([MTG], 19)	48.5	−60.5	10.5
Cuneus (20)	15.5	−91.5	22.5
Right middle occipital gyrus ([R MOG], 21)	38.5	−73.5	6.5
Fusiform gyrus (22)	29.5	−42.5	−12.5
Inferior occipital gyrus ([IOG], 23)	−36.5	−76.5	−4.5
Lingual gyrus ([LingualG], 24)	−8.5	−81.5	−4.5
Middle temporal gyrus ([MTG], 25)	−44.5	−57.5	−7.5
Cognitive-control network (CC)
Inferior parietal lobule ([IPL], 26)	45.5	−61.5	43.5
Insula (27)	−30.5	22.5	−3.5
Superior medial frontal gyrus ([SMFG], 28)	−0.5	50.5	29.5
Inferior frontal gyrus ([IFG], 29)	−48.5	34.5	−0.5
Right inferior frontal gyrus ([R IFG], 30)	53.5	22.5	13.5
Middle frontal gyrus ([MiFG], 31)	−41.5	19.5	26.5
Inferior parietal lobule ([IPL], 32)	−53.5	−49.5	43.5
Right inferior parietal lobue ([R IPL], 33)	44.5	−34.5	46.5
Supplementary motor area ([SMA], 34)	−6.5	13.5	64.5
Superior frontal gyrus ([SFG], 35)	−24.5	26.5	49.5
Middle frontal gyrus ([MiFG], 36)	30.5	41.5	28.5
Hippocampus ([HiPP], 37)	23.5	−9.5	−16.5
Left inferior parietal lobule ([L IPL], 38)	−47.5	5.5	22.5
Middle cingulate cortex ([MCC], 39)	−15.5	20.5	37.5
Inferior frontal gyrus ([IFG], 40)	39.5	44.5	−0.5
Middle frontal gyrus ([MiFG], 41)	−26.5	47.5	5.5
Hippocampus ([HiPP], 42)	−24.5	−36.5	1.5
Default-mode network (DM)
Precuneus (43)	−8.5	−66.5	35.5
Precuneus (44)	−12.5	−54.5	14.5
Anterior cingulate cortex ([ACC], 45)	−2.5	35.5	2.5
Posterior cingulate cortex ([PCC], 46)	−5.5	−28.5	26.5
Anterior cingulate cortex ([ACC], 47)	−9.5	46.5	−10.5
Precuneus (48)	−0.5	−48.5	49.5
Posterior cingulate cortex ([PCC], 49)	−2.5	54.5	31.5
Cerebellar network (CB)
Cerebellum ([CB], 50)	−30.5	−54.5	−42.5
Cerebellum ([CB], 51)	−32.5	−79.5	−37.5
Cerebellum ([CB], 52)	20.5	−48.5	−40.5
Cerebellum ([CB], 53)	30.5	−63.5	−40.5

2.2. Classification approach

Classification input data was based on loading parameter values (e.g. values that denote how well a component is expressed in subjects) generated from sc-ICA. These loading parameters were generated using a set of networks/SPECT components derived from large scale fMRI datasets. A Python-based classification program, Polyssifier (https://pypi.org/project/polyssifier/), was used to run the classification analysis. MATLAB (Version 2020b) (Matlab, 2012) was utilized to generate initial ICA figures.

53 SPECT components were estimated via the use of sc-ICA using the NeuroMark fMRI 1.0 template (Du et al., 2020a, Du et al., 2020b, Table 2). A composite spatial map denoting the 15 components that were significant after False Discovery Rate (FDR) correction from our initial analysis (Benjamini and Hochberg, 1995; Harikumar et al., 2025) is shown below in Fig. 1. We fed all 53 components into the Polyssifier classification analysis, and labels were created to distinguish patients and controls (e.g. categorical coding of ‘1's for patients and ‘0's for controls to separate the two groups). Our ground truth was to identify significant components out of the 53 between case/controls, and if the results would replicate the 15 components associated across the auditory, subcortical and sensorimotor networks in our initial study.

Fig. 1.

A spatial map from Harikumar et al., 2025) (anterior to posterior view with left and right hemispheres of the brain depicted, respectively) of corrected results (q< 0.05) with 15 loading parameter SPECT components showing components for patients < controls, with increased loading parameter functional covariance in the frontal and medial regions of the brain.

For classification, models were trained using a k-fold cross validation scheme (k = 10). This default value was selected based on the highest k-fold value from two sci-kit packages, specifically SelectKBest (Pedregosa et al., n.d.; https://bit.ly/SelectKBest) and f_regression (Pedregosa et al., n.d.; https://bit.ly/f_regression). Finally, for model evaluation, the area under the receiving operating characteristic curve (AUC) was used to calculate the cross-validated classifier score (Rakotomamonjy, 2004). In this initial work, we focused on a linear SVM approach as the main classifier of interest given SVM has been touted as a useful algorithm for identifying schizophrenia case/control differences in fMRI data. (Steardo et al., 2020).

Next, the linear SVM scores and salient features were plotted onto respective bar plots showing feature importance and feature rank. Additionally, training and testing AUC scores were calculated for eight additional classifiers including multilayer perceptron (with 50, 25 hidden layer sizes) across iterations, nearest neighbors, SVM, decision tree, random forest (10 estimators), logistic regression, naïve bayes, and voting classifiers.

3. Results

3.1. Initial linear SVM classifier results

After running Polyssifier, our goal was to determine which of the 53 loading parameter SPECT components were selected as the most salient features, or networks of interest. The classifier results were displayed as plots. The features denoted as the most salient independent components (IC) were plotted in order of importance and feature rank. This allowed us to see which SPECT components were most relevant, and broadly, which ICNs were associated with these components/features. The 15 SPECT components which survived FDR were initially plotted onto a heat map (Fig. 2a) which denoted ICNs that were correlated with each other (Harikumar et al., 2025)Our initial results from our prior study (Fig. 2a) found that the CC-SC, CC-AUD, SC-SM, and DM-AUD networks demonstrated strong covariation for controls-patient differences based on loading parameter values (e.g. values that denoted how strongly a component was expressed for each of the 7 ICNs, and more specifically across the 53 ICs). Feature indices were plotted first (Fig. 2b) along with feature rank (Fig. 2b; bottom figure). The linear SVM classification models resulted in a mean of 0.83 (SD = 0.12) (Table 3a), which we decided to focus on given the utility of SVM in prior fMRI studies (Steardo et al., 2020). Results also showed that ICs 18, 2, and 3 (middle occipital, subthalamus, and putamen) were ranked highest. These results diverged slightly from our initial study, identifying more of the visual and cognitive control networks as opposed to any auditory components. Finally, a confusion matrix and a table of sensitivity, specificity and accuracy measures across classifiers (Table 3b) further demonstrated the robust performance of linear SVM in comparison to all other classification methods.

Fig. 2a.

NeuroMark FNC map (adapted from Harikumar et al., 2025) of healthy controls-patient differences with unthresholded (lower part of the FNC triangle) and FDR corrected resting state results (upper part of the FNC triangle) showed stronger covariation with CC-SC, CC-AUD, SC-SM, and DM-AUD networks. These initial results were validated by the Linear SVM model below.

Fig. 2b.

(Top) demonstrates feature importance in order of most to least important using the linear SVM classification approach. Overall, SPECT components 18, 2, and 3 (belonging to the visual and cognitive control networks, respectively) were found to be the most important feature indices when organized by rank. Fig. 2b(Bottom) demonstrates results from a significantly better performing classifier (logistic regression), which predicted similar network initial feature results across the visual and cognitive networks.

Table 3a.

Initial classification results generated from Polyssifier for the linear SVM classification. Results show that overall, linear SVM from the testing dataset were able to classify the salient SPECT components with 0.83 AUC, and identify three ICNs of interest: auditory, subcortical, and sensorimotor networks.

Classifier Approach	Mean	Std Deviation	Minimum Value	Maximum Value
Linear SVM: Train	0.94	0.01	0.92	0.98
Linear SVM: Test	0.83	0.12	0.61	1.0

Table 3b.

A confusion matrix and subsequent results table demonstrates the sensitivity, specificity, and accuracy metrics (denoted as true negative, false positive, false negative, and true positive values) for the overall classifier performance. Overall, logistic regression outperformed all other classifiers, and demonstrated improved sensitivity and specificity, with decreased false negative performance.

Classifier	TN	FP	FN	TP
Multilayer Perceptron	53	23	25	112
Nearest Neighbors	52	24	25	112
SVM	59	17	37	100
Linear SVM	59	17	34	103
Decision Tree	44	32	38	99
Random Forest	46	30	28	109
Logistic Regression	44	32	20	117
Naïve Bayes	53	23	35	102
Voting	59	17	32	105

CLF	TN	FP	FN	TP	Sensitivity	Specificity
Multilayer Perceptron	53	23	25	112	0.8175	0.6973
Nearest Neighbors	52	24	25	112	0.8175	0.6842
SVM	59	17	37	100	0.7299	0.7763
Linear SVM	59	17	34	103	0.7518	0.7763
Decision Tree	44	32	38	99	0.7226	0.5789
Random Forest	46	30	28	109	0.7956	0.6052
Logistic Regression	44	32	20	117	0.8540	0.5789
Naïve Bayes	53	23	35	102	0.7445	0.6973
Voting	59	17	32	105	0.7664	0.7763

3.2. Additional post-hoc classifier results

Next we evaluated the performance of 8 additional classifiers to evaluate the sensitive of the result to the choice of classifier. We noted several well performing classifiers were equivalent to an AUC value of 0.85 (logistic regression), as well as nearest neighbors (0.81). By contrast, linear SVM performed more poorly (AUC = 0.75), followed by SVM (AUC = 0.72). In comparison to other classifiers, linear SVM performed poorer with decreased sensitivity and increased presence of false negatives (i.e. the model missing true relationship between networks).

4. Discussion

Classification results from this study noted particularly that after running linear SVM and logistic regression (Fig. 2b), SPECT components from the visual and cognitive control networks showed up as the most salient and highly predictive networks for patients, particularly more robustly in logistic regression vs. linear SVM. This was in comparison to the other 7 classifier approaches, which did not perform as well compared to logistic regression. It is also important to note that SPECT perfusion or cerebral blood flow is well correlated with blood flow measured by fMRI (Ernst et al., 1999) with a recent review indicating that certain patterns of hypo or hyper-perfusion relate to the manifestation of clinical symptoms in schizophrenia (Percie Du Sert et al., 2023). Taken together, fMRI and SPECT comparably measure blood flow patterns which impact the manifestation of clinical characteristics and behavioral symptoms which warrant further research. Furthermore, the identification of visual and cognitive control networks being the most salient features is unsurprising. Several studies have indicated aberrant network connectivity of the cognitive control and visual networks in schizophrenia (Harikumar et al., 2025; K. Jensen, 2026; Ray et al., 2017), with recent studies noting peripheral inflammation possibly mediating cognitive deficits across the hippocampal-thalamo-visual circuitry (Gou et al., 2026). The role of inflammation has played an important role in identifying disrupted mechanisms in schizophrenia, including aberrant face and emotion processing (Harikumar et al., 2026) which warrants further examination. It could be that complex physiological and biological processes affect visual and cognitive processing in a cascading manner. Indeed, prior studies have indicated cortico-striatal abnormalities that lead to hallucination and delusion symptoms in schizophrenia (K. M. Jensen et al., 2024). Given the limited literature translating these findings in SPECT imaging, future studies should examine the role of inflammation on functional connectivity disruptions in greater detail.

We note that strong visualizations (i.e. clear, easy to read FNC plots with warm and cool colors) help in interpreting classifier results, especially mapping the results to a clinical context. The FNC figure from our prior study (Fig. 2a) shows evidence of both strong and weak ICNs such as DM, CC, AUD and SC network connectivity that may play a direct role in positive and negative symptom expression in schizophrenia. We also note that strong and weak ICN connectivity is clearly observable with strong connectivity indicated in orange/red hues, and weak connectivity noted in blue hues; thus, clear visualizations of classifier results can assist researchers with examining these networks in greater detail for future research. Because of the clear FNC maps coupled with the classifier results, we now have a better picture of how SPECT results may map well onto prior fMRI identified networks, and possible links to clinical symptomatology. Rather than using FNC features, the loading parameters were utilized to provide a more targeted look into dysconnectivity patterns. Prior studies (Premi et al., 2017) adopting a similar approach using source based morphometry (SBM) in SPECT have noted that utilizing loading parameters provide a more robust approach in identifying specific targeted network features, as opposed to FNC alone. Thus, this approach was utilized for this study to better determine classification performance.

Finally, we observed that several classifiers performed comparably to the logistic regression approach. It could be that multiple approaches, not just logistic regression, may yield favorable results. Possible reasons for improved performance in logistic regression could be due to several reasons,

Other aspects to consider are the complexity and heterogeneity of mental disorders which may not fully be captured by these classifier approaches. Previous studies which have utilized classification approaches on SPECT datasets noted that heterogeneity among patient data add a layer of complexity to interpreting the data (Vázquez-Abad et al., 2020; Verma et al., 2023). However, our study adds an additional layer of analysis utilizing a robust spatially constrained ICA approach, thereby constraining the ICNs and ICs to the NeuroMark template (a well validated scICA template across fMRI studies).

5. Conclusion/future directions

These results provide promising initial findings for focusing on clinically relevant predictions, which could extend to various clinical interventions such as measuring medication response, risk for future psychotic episodes, and pinpointing specific brain domains associated with debilitating hallucination and delusion symptoms.

We note for the first time a scICA based approach that initially identified three ICNs of interest in schizophrenia, which surprisingly found differing results compared to our initial study in terms of the networks identified. Additionally, the pathophysiology of schizophrenia has been extensively explored through the conceptualization of dysconnectivity models in schizophrenia. Schizophrenia was discussed initially as a “disconnection syndrome” (Friston, 2002) and later refined through conceptualizations that expanded upon various neurotransmitter deficiencies (i.e. excitatory/inhibitory neurotransmitter pattern disruptions), neurodevelopmental disruptions (i.e. trauma, infection in utero), and wide network patterns of dysconnectivity across large scale brain networks noted in schizophrenia (Harikumar et al., 2023). Recent studies have expanded even further upon these theories, noting that PET and SPECT studies have further assisted in understanding neurobiological mechanisms of schizophrenia (Mohammadgholi and Hosseinimehr, 2026) due to the specific tracers used. Unlike fMRI, SPECT/PET imaging utilizes specific radiotracers which uniquely capture and understand the effects of antipsychotic medication on brain function, and it's effect on the dopaminergic system (Mohammadgholi and Hosseinimehr, 2026). Further insights into treatment resistant schizophrenia, for example, have also highlighted the utility of PET/fMRI and other modalities to explore neurotransmitter activity and structural brain changes in schizophrenia (Petch et al., 2026). Lastly, increased interest in factors such as suicidality, trauma, and inflammation along with patterns in schizophrenia subtypes have noted increased clinical symptoms associated with dysconnectivity across various social cognition and face/emotion processing networks in prior rest and task fMRI studies (Harikumar et al., 2026). Large scale FNC studies have noted relationships between medication dosage, age of onset, chronicity, symptom severity, cognitive performance and FNC patterns across sensory, motor, and cortical circuitry (K. Jensen, 2026). Taken together, these studies examine the pathophysiology of schizophrenia utilizing a wide array of techniques, methodologies, and approaches that touch on various networks including the auditory, subcortical, and sensorimotor networks which should be explored in further detail.

We also noted future considerations such as increasing the overall sample size, replication of this analysis with similar classifying approaches with a larger sample size, and careful consideration of diagnostic and patient heterogeneity. Some limitations with the existing dataset include an imbalanced sample set between case/controls. In our original study utilizing the NeuroMark 1.0 template (Harikumar et al., 2025), we addressed this limitation by performing a site comparison analysis given that site was a significant covariate in our initial regression. Specifically, we compared Site 2 (the common site between controls and patients) vs. group differences across the remaining sites. We tested the site analysis further by comparing Site 2 vs. all other site performance in a recent preprint utilizing the same dataset presented here but with a novel NeuroMark SPECT 68 component template (Harikumar et al., 2026). The group differences between healthy – patients were strongly correlated between Site 2 vs. all sites (r = 0.8472), and provide further support that site was not a confounding factor for the results, irrespective of SPECT template differences.

In the present study, logistic regression consistently performed well compared to most other classifiers irrespective of site. However, we recognize that future studies should expand the recruitment of healthy controls for a more balanced comparison between case/controls. Additionally, intake of anti-psychotic medication is an important consideration particularly regarding the influence of medication on SPECT perfusion patterns. We addressed this as well in our initial study (Harikumar et al., 2025), noting no significant group differences between a sample of matched medication/non-medication groups used as a comparative follow up analysis. We have also provided a detailed summary of medication status, antipsychotic usage, and clinical information for patients in the present study as a supplementary table.

Taken together, these results allow us and the field to consider the value of using a scICA approach and also an initial linear SVM approach to understand broad ICN patterns. Using a combined approach of scICA + classification, we identified and highlighted relevant components related to the three broader NeuroMark networks associated with dysconnectivity in schizophrenia. Replicating findings from prior fMRI studies, particularly noting these three networks of interest in schizophrenia prove promising for clinical interventions.

Our growing knowledge of dysconnectivity across visual and cognitive control networks in schizophrenia could assist in robust clinical interventions for hallucinations and cognitive dysfunction. Indeed, efforts are being made in these areas with clinical interventions such as acceptance and commitment therapy (ACT) being used to manage auditory hallucinations in patients (Ashry, 2021), along with interventions such as social skills training, metacognitive training, and cognitive remediation to manage disruptive cognitive symptoms (Barlati et al., 2024). We hope to continue this line of work with other comorbid datasets, and possibly examine how classifiers can also predict clinical symptomatology in schizophrenia. Furthermore, incorporating functional connectivity analysis and/or multimodal fusion between functional and structural connectivity (i.e. fMRI-SPECT studies) could serve as a potential starting point for future work. In conclusion, sc-ICA is a robust method for analyzing SPECT data, and classifying approaches such as linear SVM serve as a powerful tool to replicate and confirm prior study findings. These findings confirm prior fMRI results, which may then assist in clinical interventions or therapies for the future.

Statement of informed consent/approval

All patients in the study gave informed consent to have their anonymous data used in future research at the time of their initial visit to the clinic (Integ Review Board #004-Amen Clinics Inc.).

The comparison dataset consisted of 76 healthy participants (HC) who were recruited under a prior study of SPECT in healthy participants and approved by Western IRB (WIRB # 20021714).

Integ Review Board (004-Amen Clinics Inc.) on 09/19/2014 and determined to be except category 4.

The healthy participants dataset was collected under a prior study approved by Western IRB (WIRB # 20021714), approved on 10/19/2009.

Declaration of generative AI and AI assisted technologies in the writing process

During the preparation of this work the primary author used Grammarly and ChatGPT in order to check content for flow and clarity, generate and revise code, as well as proofread the work for grammar and spelling errors. After using this tool/service, the primary author reviewed and edited the content as needed and take(s) full responsibility for the content of the published article.

Funding sources

This project was funded generously by the following grant: NIH R01MH123610 (PI: Calhoun). Additionally, Amritha Harikumar held a graduate fellowship with Georgia State University including the GSU Brains and Behavior Fellowship. She is presently funded by the Amen Clinics/TReNDS Center in 2026.

CRediT authorship contribution statement

Amritha Harikumar: Data curation, Formal analysis, Investigation, Methodology, Project administration, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Joanne Wardell: Methodology, Resources, Software, Writing – review & editing. David Keator: Data curation, Methodology, Project administration, Resources, Software, Writing – review & editing. Daniel Amen: Data curation, Funding acquisition, Methodology, Project administration, Resources, Writing – review & editing. Sergey M. Plis: Resources, Software, Visualization, Writing – review & editing. Vince D. Calhoun: Conceptualization, Data curation, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Writing – review & editing.

Declaration of competing interest

The imaging data was collected by the Amen Clinics as part of their routine patient intake for clinical treatment. Upon consent (Integ Review Board (004-Amen Clinics Inc.)), patients provide their de-identified data for research use, made available by the non-profit Change Your Brain Foundation to academic and research institutions.

TReNDS used the retrospective imaging data from Change Your Brain Foundation to study functional connectivity in schizophrenia across imaging modalities.

Dr. Keator is currently the chief research officer at the Amen Clinic, and receives a salary from the clinic to perform neuroimaging research. Dr. Amen is the founder and CEO of Amen Clinics, receives salary, and owns private stock of the company.

Appendix A. Supplementary data

The following are the Supplementary data to this article:

Data availability

Data will be made available on request.