Deep Learning-based Brain Age Prediction in Patients With Schizophrenia Spectrum Disorders

Abstract

Background and Hypothesis

The brain-predicted age difference (brain-PAD) may serve as a biomarker for neurodegeneration. We investigated the brain-PAD in patients with schizophrenia (SCZ), first-episode schizophrenia spectrum disorders (FE-SSDs), and treatment-resistant schizophrenia (TRS) using structural magnetic resonance imaging (sMRI).

Study Design

We employed a convolutional network-based regression (SFCNR), and compared its performance with models based on three machine learning (ML) algorithms. We pretrained the SFCNR with sMRI data of 7590 healthy controls (HCs) selected from the UK Biobank. The parameters of the pretrained model were transferred to the next training phase with a new set of HCs (n = 541). The brain-PAD was analyzed in independent HCs (n = 209) and patients (n = 233). Correlations between the brain-PAD and clinical measures were investigated.

Study Results

The SFCNR model outperformed three commonly used ML models. Advanced brain aging was observed in patients with SCZ, FE-SSDs, and TRS compared to HCs. A significant difference in brain-PAD was observed between FE-SSDs and TRS with ridge regression but not with the SFCNR model. Chlorpromazine equivalent dose and cognitive function were correlated with the brain-PAD in SCZ and FE-SSDs.

Conclusions

Our findings indicate that there is advanced brain aging in patients with SCZ and higher brain-PAD in SCZ can be used as a surrogate marker for cognitive dysfunction. These findings warrant further investigations on the causes of advanced brain age in SCZ. In addition, possible psychosocial and pharmacological interventions targeting brain health should be considered in early-stage SCZ patients with advanced brain age.

Introduction

Over the past decade, several neuroimaging studies have proposed a biomarker of individual brain health called “brain age.” Reliable biomarkers of brain aging may have significant neuroscientific and clinical uses. Previous studies have demonstrated that schizophrenia (SCZ) is associated with accelerated brain aging compared to the chronological age estimated from brain images.^¹ To date, 20 studies have investigated brain aging in SCZ patients using machine learning (ML) algorithms applied to structural magnetic resonance imaging or diffusion tensor imaging.^1–20 SCZ patients have a higher brain-predicted age difference (brain-PAD) than healthy controls (HCs), with scores varying from 2.6^6,14 to 7.8¹⁶ years. Notably, previous studies included a relatively small number of patients (range: 43–341). Two recent large studies analyzed 1110 and 2803 SCZ patients, respectively, and demonstrated a moderate increase in brain-PAD.^4,7 Although previous studies have included heterogeneous SCZ populations, such as first-episode (FE) psychosis/SCZ and recent-onset or recurrent SCZ, no study has included patients with treatment-resistant schizophrenia (TRS). Brain-PAD had a significant negative relationship with a composite measure of cognitive function in HCs.^21,22 Only one previous study showed an association of advanced brain age with lower cognitive function in psychosis⁵; no previous study has evaluated this association in TRS patients.

Deep learning (DL) has achieved tremendous success in various tasks from computer vision and natural language processing to medical image analysis. In addition to multivariate pattern recognition and making inferences at an individual level, DL has significant advantages over other ML techniques because features are not manually extracted and are learned from the data, resulting in a more objective process with less bias. Therefore, DL techniques have promising applications in various neuropsychiatric disorders. Several methods have been used to differentiate SCZ patients and HCs.^23–29 However, to the best of our knowledge, only two studies have applied DL to predict brain age in the SCZ: cascade neural network³ and inception-resnetv2 framework.² Among the various DL methods, we are particularly interested in the simple fully convolutional network (SFCN), a new deep convolutional neural network model.³⁰ Compared to other deep network architectures, SFCN has fewer parameters and is more suitable for the analysis of small datasets and three-dimensional volume data.

The present study measured the brain-PAD in patients with schizophrenia spectrum disorders (SSDs) using SFCN regression (SFCNR) and three ML models with structural magnetic resonance imaging (sMRI). Subgroup analyses were conducted for FE-SSDs vs TRS subgroups and drug-naïve/drug-free vs medicated SCZ/FE-SSD subgroups. ML models in the present study were built based on traditional ML algorithms and features measured from sMRI using the FreeSurfer. The employed traditional ML algorithms were ridge regression (RR), support vector regression (SVR), and relevance vector regression (RVR). The features used for ML models included 7 subcortical volumes, 34 cortical thickness and 34 surface area, lateral ventricles and total intracranial volume. We chose these 77 features because they can be generalized well to independent data promoting brain age model deployability and shareability, and were used in the developed model by the Enhancing NeuroImaging Genetics through Meta-Analysis (ENIGMA) consortium. In addition, the associations among brain-PAD, cognitive function, and other clinical parameters were investigated.

Methods and Materials

Sample

For pretraining, we used T1 data of 49 123 HCs from the UK Biobank.³¹ To ensure a balanced age distribution, we randomly selected 230 HCs at each age between 48 and 80 years, resulting in a final sample size of 7590. For fine-tuning, we used T1 data from 775 healthy Asians. Among them, 288 and 209 of whom were recruited from Jeonbuk National University Hospital (JBUH) and Korea University Anam Hospital (KUAH), respectively. Data from the remaining 278 individuals were downloaded from an open science resource from the Consortium for Reliability and Reproducibility (CoRR, http://fcon_1000.projects.nitrc.org/indi/CoRR/html/index.html). Within the CoRR, we only downloaded data of Asian HCs aged 18–75 years from Beijing Normal University (n = 56), Institute of Psychology, Chinese Academy of Sciences (n = 100), Jinling Hospital Nanjing University (n = 30), Southwest University (n = 70), and Xuanwu Hospital, Capital University of Medical Sciences (n = 22).

An independent test set comprising data from 233 SSD patients was used. The patients were recruited from JBUH and diagnosed according to the DSM-IV criteria^32,33 using the Structured Clinical Interview for DSM-IV Axis I Disorders/Patient Edition.³⁴ Of the 233 patients, 196 had SCZ and 37 had FE schizophreniform disorder or schizoaffective disorder. Of the 196 SCZ patients, 35 and 46 had FE-SCZ and TRS, respectively, whereas the remaining 115 patients had chronic SCZ. The first episode was defined as an illness duration ≤2 years. Treatment resistance was defined as failure to respond to an at least 6-week trial of at least two antipsychotic medications administered in adequate doses (equivalent to at least 600 mg/day of chlorpromazine [CPZ]) and persistence of clinically relevant positive or negative symptoms (at least one positive or negative symptom with a Positive and Negative Syndrome Scale [PANSS^35,36] score ≥4). The second criterion was not applied to patients using clozapine.

HCs were recruited through advertisements and interviewed using the screening module of the Structured Clinical Interview for DSM Non-Patient Edition³³ for Axis I diagnoses. Exclusion criteria applied to both training and independent test sets were as follows: (a) no history of previous or current psychiatric disorder, neurological disorder, or significant medical condition and (b) no first-degree relatives with psychiatric disorders. All participants were aged 18–72 years. Handedness was assessed using the Edinburgh Handedness Inventory.³⁷ The participants were enrolled in the study voluntarily and provided written informed consent. The study protocol was approved by the Ethics Committees of JBUH (approval no. CUH 2019-03-046) and KUAH (approval nos. BNC 2015-AN-0009, SDPC 2015-AN-0073, and DTIC 2018-AN-01118), and the study was conducted in accordance with the Declaration of Helsinki.

Demographic and Clinical Parameters

We collected sociodemographic information for Asian HCs and patients, which included age, sex, handedness, and years of education. However, the years of education were not available from the CoRR. Symptom severity in patients was evaluated within 1 week of performing functional magnetic resonance imaging using PANSS^35,36 and the Social and Occupational Functioning Assessment Scale.³⁸ In addition, cognitive function was assessed using computerized neurocognitive tests (MaxMedica, Inc., Seoul, Korea). The cognitive domains of attention, verbal memory, executive functioning, and language were evaluated using the digit span assessment and auditory continuous performance tests, a verbal learning test, the Wisconsin Card Sorting Test, and a word fluency test, respectively. Composite Z-scores for each cognitive domain were calculated using the mean and standard deviation of HCs. We reversed the commission and perseverative error scores so that positive Z-scores indicated better performance. Global cognitive function was estimated by averaging the Z-scores of the four cognitive domains. The results are shown in Supplementary table S1.

Image Acquisition, Preprocessing, and Analysis

Structural T1-weighted scans of participants were acquired at JBUH or KUAH or downloaded from the CoRR. Cortical parcellation was performed based on the Desikan/Killiany atlas.³⁹ FreeSurfer was used to segment 14 subcortical gray matter regions (nucleus accumbens, amygdala, caudate, hippocampus, pallidum, putamen, and thalamus), two lateral ventricles, the thickness of 68 cortices, 68 surface area measures, and the total intracranial volume. The segments were visually inspected and statistically examined for outliers. Further details on the dataset (sample size, image acquisition parameters, and quality control) are presented in Supplementary table S2. Data were pre-processed using FreeSurfer to calculate the input values in the ML regression models.

For the SFCNR model, we pre-processed T1 data using the DL-based brain extraction toolbox HD-BET⁴⁰ to remove skull and non-brain areas. Next, we conducted linear registration using the FSL-FMRIB’s Linear Image Registration Tool^41,42 to align the individual brain space with 1-mm standard Montreal Neurological Institute space. To address variations among scanners, we applied harmonization to remove site effects and preserved biological covariates using the ComBat harmonization software.⁴³ Due to resource limitations, we used average pooling with kernel and stride size of 2, and conducted min-max normalization to scale data between 0 and 1.

ML Models

To estimate the normative brain age models, we combined the FreeSurfer measures from the left and right hemispheres by calculating the mean ([left + right]/2) volumes of subcortical regions and lateral ventricles, as well as cortical region thickness and surface area, resulting in 77 features. Because data were collected from different institutes, scanner effects were removed using ComBat,^43,44 which assumes a unique linear model of location and scale at each feature and that scanners (or sites) have both additive and multiplicative effects on data. To verify the removal of scanner effects, we performed principal component and linear discriminant analyses (Supplementary figure S1).

We used HC datasets from JBUH and CoRR as training samples and the HC dataset from KUAH and patient dataset from JBUH as independent test samples (figure 1). Principal component analysis was performed to check for outliers in the training HC dataset and 25 samples were removed (Supplementary figure S2). Using RR, SVR, and RVR, and the HC training samples, we estimated the normative models for the association between the 77 average structural brain measures and age. We used the “glmnet_4.1-2” package for RR and the “kernlab_0.9-29” package for SVR and RVR. During training, a nested 10-fold cross-validation scheme was used to search for the optimal tuning parameter lambda for RR and the sigma of kernel function for SVR and RVR in the inner loop, as well as to evaluate the predictive performance of the model trained on the outer test set using the best inner trained model with the minimum mean absolute error (MAE). We selected the best brain age predictor from the 10 outer models based on the minimum MAE and trained it on the entire dataset to obtain the final brain age estimator.^45–47 To assess model performance using the training samples, we calculated MAE, Pearson’s correlation coefficient (r) between predicted brain age and chronological age, and the proportion of the variance explained by the model (R²).

Fig. 1.

Schematic illustration of data partition into training and test samples and development of brain age prediction model. FE-SSDs, First-Episode Schizophrenia Spectrum Disorders; RR, ridge regression; RVR, relevance vector regression; SCZ, schizophrenia; SVR, support vector regression; TRS, treatment-resistant schizophrenia; PCA, principal component analysis; ^†Jeonbuk National University Hospital (JBUH) (n = 288), Consortium for Reliability and Reproducibility (CoRR) (n = 278) and Korea University Anam Hospital (KUAH) (n = 209); ^††JBUH (n = 267) and CoRR (n = 274); ^§KUAH (n = 209).

SFCNR Model

We constructed a convolutional neural network-based regression model as the backbone network using SFCN³⁰ consisting of six blocks. Details about SFCN are described in the Supplementary Figure S3. The first five blocks were composed of the 3D convolutional layer with a kernel size of 3, batch normalization, max pooling with a kernel size 2, and rectified linear unit (ReLU) activation function.⁴⁸ The sixth block consisted of a 3D convolutional network with a kernel size of 1, batch normalization, and ReLU activation function. Following the backbone network, a seventh block was added, which consisted of a flattening layer, linear layer, ReLU activation function, 50% random drop out layer, and a linear layer for the regression.

We pretrained the model with selected data (n = 7590) from UK Biobank (https://biobank.ctsu.ox.ac.uk).⁴⁹ We applied two augmentation methods to these pretraining data for the training set, ie, shifting every axis for a randomly selected voxel size of 0–2 and mirroring along the sagittal axis.³⁰ The data were split at an 8:1:1 ratio into training, validation, and test sets, and trained with the MAE loss between chronological age and predicted age. We utilized 16 batch sizes for 300 epochs, and applied early stopping when the validation loss did not decrease for 100 epochs. A stochastic gradient descent optimizer⁵⁰ with a learning rate l of 0.01 and weight decay of 0.001 was used. In addition, the stepwise learning rate decay was calculated for every step size of 30 with a learning weight γ of 0.3 according to the backbone network. The pretraining learned parameters were applied to the data from 541 Asian HCs to improve brain age prediction performance. Fine-tuning was performed by selecting optimized hyper-parameters based on the nested cross-validation. We used 16 batch sizes for 50 epochs. An adaptive moment estimation optimizer⁵¹ with a learning rate of 5e−4 and 1e−6 weight decay was used.

Model Validation

Model performance was validated in the independent test samples of controls and patients. The parameters learned from training on the controls were applied to the test samples of controls and patients to obtain the brain-based age estimates. To assess model performance using these test samples, we also calculated the MAE, r, and R². To evaluate the generalizability and independently test the model on HC samples acquired from independent scanning sites, we applied the trained model to HCs from KUAH. The brain age was also estimated for patients. Then, brain-PAD (predicted brain-based age − chronological age) was calculated for each individual.

Considering the significant correlation between chronological age and brain PAD⁵² and nonlinear age effects,⁵³ we corrected for the two effects post hoc by adding age or age² as a covariate to the statistical models, which were Brain-PAD = intercept + β1(sex) + β2(age) and [+ β3(age²)] + ε. We compared the linear and quadratic models using the goodness-of-fit test and selected the final models, which were a quadratic model for RR and SVR, and a linear model for RVR and SFCNR. The significant relationship between chronological age and brain-PAD disappeared after controlling for age, age,² and sex (Supplementary figures S4-1, S5-1, S6-1, and S7-1).

Statistical Analyses

The demographic and clinical characteristics of the training and test samples were compared using analysis of variance (ANOVA) and the chi-square test. Statistical analyses for brain-PAD were conducted in the test samples only. To compare the corrected brain-PAD among the groups, we performed Student’s t-test or ANOVA with Bonferroni correction. Correlation analysis was performed between the brain-PAD and clinical parameters among patients. False discovery rate adjustments were made for multiple comparisons.

Results

Comparison of Demographic and Clinical Characteristics Among the Training and Test Sets

Significant differences in age (P < .001), sex (P = .001), and education (P = .007) were observed among the training HCs, test HCs, and test SCZ patients. Age (P < .001) and sex (P = .002) were significantly different among the test HCs, test FE-SSD patients, and test TRS patients (Supplementary table S3). The age distributions in the training and test datasets are shown in figure 2E.

Fig. 2.

Distribution of MAEs by RR (A), SVR (B), RVR (C), and SFCNR (D) and Age Distribution (E) among training and test samples. FE-SSDs, First-Episode Schizophrenia Spectrum Disorders; HC, healthy control; RR, ridge regression; RVR, relevance vector regression; SCZ, schizophrenia; SFCNR, simple fully convolutional network regression; SVR, support vector regression; TRS, treatment-resistant schizophrenia.

Model Performance and Brain-PAD in the Test Samples

SFCNR exhibited the lowest MAE, and highest r and R² values, in the training and test datasets, whereas, RVR exhibited the highest MAE value and lowest r and R² values in the training and test datasets (table 1). The corrected brain-PADs calculated from SFCNR were −1.62 ± 6.63, 1.47 ± 8.59, 1.38 ± 7.51, and 4.06 ± 9.12 for test HCs, test SCZ patients, test FE-SSD patients, and test TRS patients, respectively (table 2). Among all four brain age prediction models, significant differences were observed between test HCs and test SCZ patients, and among test HCs, test FE-SSD patients, and test TRS patients. A post hoc test of the latter revealed significant differences in all pairwise comparisons of RR, SVR, and SFCNR, except for the comparison of test FE-SSD patients with test TRS patients in RVR, SVR, and SFCNR (table 2). Density plots showed case–control differences in brain-PAD, although large within-group variations and between-group overlap were observed (Supplementary figures S4-2, S5-2, S6-2, and S7-2 for SVR, RR, RVR, and SFCNR, respectively).

Table 1.

Brain Age Prediction Performance on Training and Test Samples

	Training HC (n = 541)			Test HC (n = 209)			Test Schizophrenia (n = 196)			Test FE-SSDs (n = 72)			Test TRS (n = 46)
	MAE	r	R 2	MAE	r	R 2	MAE	r	R 2	MAE	r	R 2	MAE	r	R 2
RR	5.79	0.85	0.72	7.24	0.79	0.63	9.01	0.68	0.46	9.55	0.55	0.30	8.60	0.71	0.51
RVR	5.79	0.69	0.47	10.08	0.57	0.32	9.35	0.53	0.28	10.88	0.33	0.11	9.79	0.32	0.10
SVR	3.44	0.93	0.87	6.72	0.82	0.67	7.78	0.64	0.41	8.82	0.54	0.29	6.85	0.61	0.37
SFCNR	1.84	0.99	0.98	5.80	0.90	0.81	6.13	0.83	0.69	4.12	0.79	0.63	7.14	0.76	0.58

Note: FE-SSDs, First-Episode Schizophrenia Spectrum Disorders; HC, healthy control; MAE, mean absolute error; r, Pearson correlation coefficients; R², proportion of the variance explained by the model; RR, ridge regression; RVR, relevance vector regression; SVR, support vector regression; SFCNR, simple fully convolutional network regression; TRS, treatment-resistant schizophrenia.

Table 2.

Brain-Predicted Age Difference^† in Test Samples Obtained Using RR, RVR, SVR, and SFCNR

	Test HC (n = 209)		Test Schizophrenia (n = 196)		P value1	Test FE-SSDs (n = 72)		Test TRS (n = 46)		P value2	1 vs 3	1 vs 4	3 vs 4
	Mean	SD	Mean	SD		Mean	SD	Mean	SD
RR	−3.09	7.57	3.11	9.65	<.001	1.85	8.66	5.77	7.32	<.001	<0.001	<0.001	0.024
RVR	−2.09	6.47	1.80	8.89	<.001	1.66	7.88	3.17	9.04	<.001	<0.001	<0.001	0.802
SVR	−2.30	6.59	2.19	8.08	<.001	1.67	8.18	4.34	6.32	<.001	<0.001	<0.001	0.127
SFCNR	−1.62	6.63	1.47	8.59	<.001	1.38	7.51	4.06	9.12	<.001	0.008	<0.001	0.149

Note: FE-SSDs, First-Episode Schizophrenia Spectrum Disorders; HC, healthy control; RR, ridge regression; RVR, relevance vector regression; SD, standard deviation; SVR, support vector regression; SFCNR, simple fully convolutional network regression; TRS, treatment-resistant schizophrenia; Test HC, test schizophrenia, test FE-SSDs and test TRS are labeled as 1, 2, 3 and 4.

^†Adjusted for age, age² and sex for RR and SVR and age and sex for RVR and SFCNR; P-value,¹t-test between 1 and 2; P-value,² ANOVA among 1, 3, and 4 and post hoc comparisons were corrected using Bonferroni test.

Correlation with Clinical Parameters

In SCZ patients, brain-PADs were significantly associated with the duration of untreated psychosis (DUP; r = −.21, P = .0269), total (r = .14, P = .047 for RR and r = .15, P = .043 for SVR), positive (r = .17, P = .021 for RVR) and negative (r = .16, P = .025 for RR and r = .17, P = .019 for SVR) symptom scores of PANSS, Social and Occupational Functioning Assessment Scale score (r = −.15, P = .032 for RR and r = −.15, P = .031 for SVR), CPZ equivalent (r = .20, P = .012 for RR, r = .28, P = .0003 for RVR, and r = .20, P = .009 for SVR), global cognitive function (r = −.19, P = .012 for RVR, and r = −.16, P = .028 for SFCNR), verbal memory (r = −.15, P = .043 for RVR), executive function (r = −.16, P = .027 for SFCNR), and language (r = −.18, P = .016 for RR, r = −.18, P = .013 for RVR, r = −.19, P = .009 for SVR, and r = −.15, P = .044 for SFCN). When adjusted for multiple comparisons, the findings on global cognitive functioning and language were remained significant (P = .033 for RVR on global cognitive functioning, and P = .033 for RVR, and P = .045 for SVR on language).

In FE-SSD patients, the CPZ equivalent (r = .38, P = .021 for RR and r = .35, P = .031 for SVR), executive function (r = .28, P = .017 for RR and r = .27, P = .021 for SVR), and language (r = −.24, P = .046 for SVR) had significant associations with brain-PAD. In TRS patients, no clinical items had significant associations with brain-PAD (table 3). Nothing remained significant after FDR correction in both FE-SSD and TRS.

Table 3.

Correlation Between Brain-PAD^† and Clinical Parameters in SCZ, FE-SSDs, and TRS

	SCZ					FE-SSDs					TRS
	n	RR	RVR	SVR	SFCNR	n	RR	RVR	SVR	SFCNR	n	RR	RVR	SVR	SFCNR
DUP	114	−0.13*	−0.12	−0.15	−0.21*	24	−0.12	0.22	0.00	0.12	46	0.23	0.07	−0.14	0.21
DI	196	0.04	0.00	−0.01	0.07	72	0.00	−0.18	−0.05	−0.01	46	0.03	−0.1	0.17	−0.12
PANSS	195					72
Total		0.14*	0.14	0.15*	0.1		0.07	0.05	0.02	0.08	46	0.12	0.06	0.22	0.09
Positive symptoms		0.13	0.17*	0.12	0.07		0.05	−0.03	−0.01	0.03	46	0.16	0.16	0.2	0.11
Negative symptoms		0.16*	0.11	0.17*	0.14		0.08	0.09	0.08	0.08	46	0.09	−0.04	0.15	0.14
General psychopathology		0.09	0.09	0.10	0.06		0.04	0.06	−0.02	0.08	46	0.09	0.07	0.22	0.01
CDSS	195	0.08	−0.02	0.05	−0.01	72	−0.05	−0.11	−0.18	0.00	46	−0.03	−0.14	0.14	0.14
SOFAS	194	−0.15*	−0.09	−0.15*	−0.12	72	0.07	−0.03	0.04	0.02	46	−0.11	−0.02	−0.20	−0.10
CPZ equivalent	163	0.20*	0.28**	0.20**	0.08	37	0.38*	0.3	0.35*	0.32	46	−0.06	0.01	0.29	−0.08
ETI Total	104	−0.15	−0.1	−0.09	0.00	21	0.00	0.16	−0.08	0.21	40	−0.24	−0.01	−0.12	−0.11
Cognitive function
Global cognitive functioning	182	−0.14	−0.19*,ǂ	−0.13	−0.16*	70	0.06	0.00	0.05	0.03	45	−0.06	−0.09	0.10	−0.2
Attention	188	−0.13	−0.13	−0.11	−0.11	72	0.00	0.06	0.07	0.07	46	−0.03	−0.03	0.07	−0.27
Verbal memory	189	−0.06	−0.15*	−0.06	−0.11	72	−0.01	−0.04	−0.1	−0.07	46	0.04	−0.06	0.21	−0.02
Executive function	191	−0.08	−0.09	−0.05	−0.16*	72	0.28*	0.07	0.27*	0.05	46	−0.17	0.05	0.04	−0.18
Language	186	−0.18*	−0.18*,ǂ	−0.19**,ǂ	−0.15*	70	−0.18	−0.17	−0.24*	−0.1	45	−0.22	−0.15	−0.01	−0.07

Note: Data given as estimate (r).

CDSS, Calgary Depression Scale for Schizophrenia; CPZ, chlorpromazine; DI, duration of illness; DUP, duration of untreated psychosis; ETI, early trauma inventory; SCZ, schizophrenia; FE-SSDs, First-Episode Schizophrenia Spectrum Disorders; HC, healthy control; PAD, predicted age difference; PANSS, Positive and Negative Syndrome Scale; SOFAS, Social and Occupational Functioning Assessment Scale; TRS, treatment-resistant schizophrenia; RR, ridge regression; RVR, relevance vector regression; SVR, support vector regression; SFCNR, simple fully convolutional network regression.

*Uncorrected P < .05; ^**uncorrected P < .01.

^†Adjusted for age, age² and sex for RR and SVR and age and sex for RVR and SFCNR.

^ǂFalse Discovery Rate-adjusted P < .05.

Discussion

There is increasing recognition of the importance of maintaining a healthy brain during aging as a societal goal. Psychiatric disorders, particularly SCZ, are associated with an increased risk of aging-related medical conditions, including cardiovascular disease, diabetes, and obesity,⁵⁴ and early mortality.⁵⁵ Using SFCNR and three commonly employed ML algorithms, we measured the brain age of SSD patients. Furthermore, we performed subgroup analyses (FE-SSD vs TRS subgroups and antipsychotic-naïve/antipsychotic-free vs medicated SCZ/FE-SSD subgroups). In addition, we examined the correlations between estimated brain age and clinical parameters. We found significantly higher brain-PAD, and significant correlations with CPZ equivalents and cognitive function, among SSD patients.

MAE is a simple metric that calculates the absolute difference between actual and predicted values, and is commonly used to assess model performance in brain-PAD. However, given the nonlinear relationships between chronological age and multimodal brain imaging features of gray matter, white matter, and functional connectivity,^56,57R² is recommended for the evaluation of model prediction performance.⁵⁸ Among training HCs, in terms of the MAE, the best model was SFCNR and the worst models were RR and RVR, regardless of the regression metric. The MAE for the training HCs using SFCNR was 1.84 years, which is significantly lower than those reported in previous studies of ML and DL algorithms (4.31 years,¹ 3.99 years,³ 6.49 years,¹⁸ and 3.702 years²). These findings suggest that the selection of learning algorithm is important.¹¹ Furthermore, automatic feature extraction and hierarchical representation learning from visual data are superior for DL vs ML algorithms with manual feature selection. The outperformance of ML by DL in predicting brain age is in line with previous studies.^59,60 However, notably, the MAE for the test HCs was 5.80 using SFCNR, which was almost 3-fold higher than that for training HCs. This suggests that the trained SFCNR model has limited generalizability. Most previous brain age prediction studies did not compare model performance among independent test HCs. In studies using Gaussian process regression and RR, the MAEs in the training and hold-out test datasets were similar (6.16 and 6.67, respectively,⁶¹ and 1.69 and 1.41, respectively⁶²). These findings raise the question of why there is a greater change in MAE for independent test HCs with the SFCNR. This may be due to overfitting of the trained SFCNR model. We suspect that the overfitting might be caused by the mismatch of age distribution or ethnical difference between the UK Biobank and Korean HC samples. This needs to be further investigated.

The brain-PAD in test SCZ patients was 1.47 ± 8.59 years using SFCNR, which was lower than in other studies (2.56 ± 6 years,¹⁴ 3.36 ± 5.87 years,¹ and 7.8 ± 1.62 years¹⁶). However, notably, about one-fourth of SCZ patients in Shahab et al study¹⁶ were older adults (mean age, 64.18 ± 8.52 years). In this study, the brain-PAD in the test FE-SSD patients was 1.38 ± 7.51 years using SFCNR, which was comparable or lower than those in previous studies (2.64 ± 4.15 years⁶ and 2.64; 95% confidence interval, 1.80–3.48⁸) of FE psychosis and SCZ. Importantly, the brain-PAD in the test TRS patients was 4.06 ± 9.12 years using SFCNR. This finding is particularly significant because it represents the first report of the brain-PAD in TRS patients. However, the SFCNR model did not exhibit a significant difference in the brain-PAD between the TRS and FE-SSD patients. This was an unexpected finding given the ample evidence that TRS patients have widespread structural abnormalities in the brain.^63,64 This suggests that the clinical utility of SFCNR may be limited, despite it being the best model in terms of the MAE for the training HC dataset. This low performance by the SFCNR model might be due to mismatch of age distribution between the training HC and TRS datasets. However, the RR model showed significantly advanced brain-PAD in TRS patients compared to FE-SSD patients. This finding is consistent with those of a previous study⁹ that demonstrated an increase in brain-PAD among at-risk, recent-onset, and recurrent SCZ patients. Furthermore, a longitudinal neuroimaging study demonstrated a progressive increase in brain age gap in SCZ patients.¹ Taken together, the results of ML algorithms suggest that after onset, there may be additional progressive pathogenic processes, despite the conceptualization of SCZ as a neurodevelopmental disorder.⁶⁵

In this study, there were no significant associations between brain-PAD obtained by the SFCNR model and DI and PANSS scores, although there was a negative association between the DUP in SCZ patients and brain-PAD predicted by SFCNR. The latter finding is counterintuitive because DUP involves active pathological processes in the brain and is often associated with poor clinical outcomes.⁶⁶ As DUP represents the sum of pathophysiological processes driven by the disease, medication, or environmental factors, DUP and brain-PAD may not overlap. Interestingly, three ML algorithms showed significant associations of brain-PAD with antipsychotic dosage (CPZ equivalent) in SCZ patients while SFCNR showed a trend toward significance. Antipsychotic-associated neuronal changes in the brain are a controversial topic in psychiatry.⁶⁷ However, accumulating evidence suggests detrimental effects of antipsychotics on brain tissue in both clinical⁶⁸ and preclinical⁶⁹ studies. Hence, we speculate that antipsychotic-induced neuronal changes contribute to advanced brain aging in patients. This inference is supported by the subgroup analysis showing that the brain-PADs of antipsychotic-naïve/free SCZ patients were significantly lower than those of medicated patients, albeit only in the RR and SVR models (Supplementary table S4). However, considering that higher and long-term exposure to antipsychotic medication is common in TRS patients, our inference does not explain the negative results of the SFCNR and all three ML models in terms of the association between brain-PAD and CPZ equivalent in TRS patients. A likely explanation is that the brain-PAD in TRS patients represents different neuroanatomical abnormalities driven by genetic factors, premorbid neurodevelopmental deficits, poor lifestyle, and cognitive function. Most importantly, the brain-PAD according to SFCNR and other ML models was associated with cognitive function in SCZ patients. Studies of brain age in healthy individuals have demonstrated that multimodal neuroimaging data can identify cognitive impairment,^21,70 suggesting that higher brain-PAD in SCZ patients can be used as a surrogate marker for cognitive dysfunction. As the brain-PAD is an index reflecting overall brain health, not just cognitive function, using the brain-PAD as a surrogate marker for cognitive function or brain health do have certain advantages over neuropsychological test. Given that TRS patients exhibit widespread structural brain changes^63,64 and cognitive impairment,^71,72 the lack of association between the brain-PAD and cognitive function was unexpected. This may be explained by the small number of TRS patients or patient characteristics; in terms of the latter factor, only stable outpatients with TRS were recruited and Z-scores of cognitive function (Supplementary table S1) were relatively low compared to previous studies.⁷³

Some limitations of the present study should be considered. First, confounding factors that can affect brain-PAD, such as lifestyle factors, behaviors, number of relapses, and cumulative antipsychotic dose, were not measured. Second, most patients were using antipsychotic medications at the time of imaging, making it difficult to differentiate the effects of medication use and illness on brain aging. Third, the age distribution of training HCs was not uniform (figure 2E). Feng et al⁷⁴ demonstrated that the MAE with the non-uniform dataset was not evenly distributed across the lifespan compared to that with balanced dataset. We reported MAE distribution of training and test samples (figure 2A–D). Future studies should include HCs with an even age distribution. Fourth, as the use of mirroring in the sagittal plane as a data augmentation when training the SFCNR model could potentially lose information on lateralized age-related brain structural changes especially in patients with SCZ, this may have affected the results. Fifth, due to the unexplainable nature of DL, it is difficult to identify the brain features that provided useful information and contributed to the prediction of brain age. Nevertheless, the strength of the present study is that this is the first to investigate staging-specific brain aging in FE-SSD and TRS patients using the DL algorithm.

In conclusion, using SFCNR and ML algorithms, we observed more advanced brain aging in patients compared to HCs, and brain aging increased in the order of FE-SSDs, SCZ, and TRS. However, differential brain-PAD between TRS and FE-SSD patients was observed using RR, but not SFCNR. This may indicate an overfitting problem of SFCNR and limited applicability in TRS patients. In SCZ and FE-SSD patients, CPZ equivalent and cognitive function were associated with the brain-PAD obtained using SFCNR and ML models. These findings warrant further investigations on the causes of advanced brain age in SCZ such as relationships between inflammatory markers or environmental factors, and advanced brain age. In addition, possible psychosocial and pharmacological interventions targeting brain health should be considered in early-stage SCZ patients with advanced brain age.

Supplementary Material

Supplementary material is available at https://academic.oup.com/schizophreniabulletin/.

Conflict of Interest

The authors report no biomedical financial interests or potential conflicts of interest.

Funding

The study was supported by Korean Mental Health Technology R&D Project, Ministry of Health and Welfare, Republic of Korea (HL19C0015), Korea Health Technology R&D Project through the Korea Health Industry Development Institute funded by the Ministry of Health and Welfare, Republic of Korea (HR18C0016), and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2019-0-00079, Artificial Intelligence Graduate School Program (Korea University)).

Author contributions

Chung YC designed the study and raised funding. Suk HI supervised the study and contributed to the final version of the manuscript. Kim WS, Heo DW, Maeng J, Shen J, Tsogt U, Odkhuu S, Zhang XF, Cheraghi S, Kim SW, Ham BJ, and Sui J acquired and analyzed the data. Kang CY and Rami FZ performed the statistical analysis. All authors discussed the findings and contributed to the final version of the manuscript.

Availability of Data Materials

The data used for pretraining SFCNR was sourced from the UK Biobank database, which is available at https://www.ukbiobank.ac.uk. As such, the investigators within the UK Biobank contributed to the design and implementation of the UK Biobank and/or provided data but did not participate in the analysis or writing of this paper. The CoRR dataset used in the study were accessed via http://fcon_1000.projects.nitrc.org/indi/CoRR/html/index.html.

Availability of Code

Our implementation source code for the deep learning method is available at https://github.com/ku-milab/SFCNR; the code and results with machine learning methods and statistical analyses are available at https://github.com/Sophy2019/Brain-age_SFCNR-ML.