Effects of Antipsychotic Medications and Illness Duration on Brain Features That Distinguish Schizophrenia Patients

Abstract

Background and Hypothesis

Previous studies have reported effects of antipsychotic treatment and illness duration on brain features. This study used a machine learning approach to examine whether these factors in aggregate impacted the utility of MRI features for differentiating individual schizophrenia patients from healthy controls.

Study Design

This case-control study used patients with never-treated first-episode schizophrenia (FES, n = 179) and long-term ill schizophrenia (LTSZ, n = 30), with follow-up of the FES group after treatment (n = 71), a group of patients who had received long-term antipsychotic treatment (n = 93) and age and sex-matched healthy controls (n = 373) for each patient group. A multiple kernel learning classifier combining both structural and functional brain features was used to discriminate individual patients from controls.

Study Results

MRI features differentiated untreated FES (0.73) and LTSZ (0.83) patients from healthy controls with moderate accuracy, but accuracy was significantly higher in antipsychotic-treated FES (0.94) and LTSZ (0.98) patients. Treatment was associated with significantly increased accuracy of case identification in both early course and long-term ill patients (both p < .001). Effects of illness duration, examined separately in treated and untreated patients, were less robust.

Conclusions

Our results demonstrate that initiation of antipsychotic treatment alters brain features in ways that further distinguish individual schizophrenia patients from healthy individuals, and have a modest effect of illness duration. Intrinsic illness-related brain alterations in untreated patients, regardless of illness duration, are not sufficiently robust for accurate identification of schizophrenia patients.

Introduction

The development of biological markers for common psychiatric disorders has remained relatively undeveloped compared to other fields of medicine.^1,2 Neuroimaging with MRI offers one of the most promising biomarker approaches in psychiatry, particularly in schizophrenia in which quantitative functional and anatomical brain alterations are relatively pronounced and well established.³ Initial evaluation of biomarker utility has focused on case-control differentiation. There has been a growing use of machine learning (ML) for this purpose, as it evaluates utility at the individual rather than group level and it considers features together simultaneously rather than in separate univariate comparisons.^4–6 Variable medication history and illness duration, and focus on single rather than multimodal MRI data, may account for variability in findings across prior studies.^7,8

Many studies have demonstrated the impact of antipsychotic medications on brain anatomy and function.^9–12 For example, short-term treatment is associated with increased functional activity, particularly in the bilateral prefrontal and parietal cortex, left superior temporal cortex, and right caudate nucleus.⁹ Structural studies have shown increased volume in the striatum and decreased volume in frontal cortex in long-term treated schizophrenia but not in untreated patients.¹³

Progression of brain changes over the course of illness has also been reported.^13–16 Several studies have shown widespread typically modest volumetric abnormalities in frontotemporal, thalamocortical, and subcortical-limbic circuits in never-treated first-episode schizophrenia patients.^15,17 Accelerated age-related cortical thinning has been reported in the right pars triangularis, right ventromedial prefrontal cortex, and left superior temporal gyrus in long-term ill never-treated patients.¹⁵ In a recent meta-analysis, long-term schizophrenia patients showed greater cortical thinning in right insula, right inferior frontal cortex, left lateral temporal cortex, and right temporal pole compared with first-episode schizophrenia patients.¹⁸

Independently examining drug treatment and illness course effects is complicated, as treatment duration ideally parallels illness duration. To address this challenge we recruited a large patient sample (n = 373) that included groups of never-treated first-episode schizophrenia patients (NT-FES), a subgroup of which were rescanned after one-year of antipsychotic treatment (AT-FES), long-term ill but never-treated schizophrenia patients (NT-LTSZ) and long-term ill schizophrenia (AT-LTSZ) patients who had an extensive antipsychotic treatment history, scanned using a single 3.0 T MR scanner and acquisition protocol. Matched controls were recruited for each of these 4 patient groups. Each of the 4 patient groups with their respective matched control samples were examined separately using Multiple kernel learning (MKL), which offers promise for combing various features relative to conventional support vector machine approaches. Classification results from the four case-control analyses were contrasted with 2 aims: First to determine how antipsychotic treatment and course of illness impact performance of multimodal MRI data in schizophrenia identification and second to identify the image features that were related to antipsychotic medication and illness duration in the different case-control comparisons.

Methods

Participants

Scans were performed between 2006 and 2017 at West China Hospital of Sichuan University. Diagnosis of schizophrenia was determined using the Structured Interview for the DSM-IV (SCID). Time from illness onset was determined using the Nottingham Onset Schedule.¹⁹ Exclusion criteria for all participants were: (1) contraindications for MRI examination, (2) any neurological disorder or head injury, (3) lifetime drug or alcohol abuse or dependency, (4) pregnancy, and (5) significant systemic illness such as cardiovascular disease. Finally, a total of 746 right-handed individuals were used. Never-treated first-episode schizophrenia patients (n = 179, illness duration: 11.3 ± 19.9 months) were studied before treatment, 71 patients of whom were rescanned after 1 year of antipsychotic treatment (n = 71, illness duration: 20.3 ± 11.4 months). Long-term ill schizophrenia patients included those who had never been treated (n = 30, illness duration: 262.7 ± 155.0 months), and a group who had a long-term history of antipsychotic treatment (n = 93, illness duration: 226.9 ± 124.4 months). The average chlorpromazine equivalent and antipsychotic types patients had taken are shown in table 1. While for long-term patients, because of the long history of treatment by different providers, we estimated average medication dosages for the most recent 10 years using all available data. Each of the 4 patient groups had its own matched sample of healthy controls. Details regarding patient samples are provided in table 1 and Supplementary table S1 (see Supplementary Materials). The study was approved by the Institutional Review Board (IRB) of West China Hospital of Sichuan University. Written informed consent was obtained from participants, and from their legal guardians for those who were under the age of 18 years.

Data Acquisition and Preprocessing

MRI data were acquired from a GE Signa EXCITE 3.0 T scanner (GE Healthcare) equipped with an 8-channel phased array head coil. All participants were scanned with a uniform protocol including axial high-resolution T1-weighted imaging and resting-state fMRI. Acquisition protocols are detailed in Supplementary Materials. Before data processing, a data quality check were carried out to determine whether data quality varied across groups. No significant group differences were found in signal-to-noise ratio or in head movement parameters for fMRI scans.

Functional brain imaging parameters including regional homogeneity (ReHo)²⁰ and fractional amplitude of low-frequency fluctuation (fALFF)²¹ were calculated via Data Processing Assistant for Resting-State fMRI software (advanced edition, DPARSFA, version 5.0, http://www.rfmri.org/DPARSF).²² Cortical thickness (CT) and cortical surface area (SA) measures were obtained using the FreeSurfer package (version 6.0.0, https://surfer.nmr.mgh.harvard.edu/). Gray matter volume (GMV) was determined using optimized voxel-based morphometry (VBM), following diffeomorphic anatomical registration through exponentiated lie algebra (DARTEL)²³ using Statistical Parametric Mapping software (SPM12, https://www.fil.ion.ucl.ac.uk/spm/). Image preprocessing details are presented in Supplementary table S2.

Features Used in Machine Learning

CT and SA were obtained from all neocortical regions of the Desikan-Killiany Atlas (including 68 cortical areas),²⁴ while GMV and functional parameters were extracted using the Anatomical Automatic Labeling (AAL) template (including 90 cerebral regions) and obtained in both cortical and subcortical gray matter regions. The 2 atlases are commonly used for anatomic and functional studies respectively, so we used them in this way to be consistent with the most widely used approaches. All brain regions for functional and anatomical measures were put into feature selection. Then we performed a random forest-based feature selection algorithm (implemented in R package, “Boruta”; https://www.rproject.org/) to select all relevant features.²⁵ Shadow features which were copies of all features were added to the original dataset and shuffled. Then the random forest classifier on the extended dataset evaluated the relevance of real features by comparing the contribution between real features and shadow features. At each iteration, the algorithm tagged the features with relevance by checking whether a real feature had a higher relevance to classification than the best of its shadow features. The algorithm was set to terminate when the relevance of all features was established. Only brain regions that were tagged as relevant features in the algorithm were put as selected features in the classification.

Multiple Kernel Learning

A sparse version of MKL was performed to integrate multi-parameter relevant features for classification using the Simple MKL algorithm, implemented in PRoNTo Toolbox (http://www.mlnl.cs.ucl.ac.uk/pronto/prtsoftware.html).²⁶ Linear kernels per modality were built from individual regional CT, SA, GMV, ReHo, and fALFF selected features. A nested cross-validation (CV) was used to perform hyper-parameter optimization (5-fold CV for the inner and outer loop). We used accuracy, sensitivity, and specificity to evaluate classification performance across the four case-control group pairs. The process was repeated 20 times independently to avoid potential bias caused by random partition of subsets. Statistical significance of classification accuracy was determined by permutation tests repeated 1000 times using a threshold of p < .001 (2-tailed). To assess the importance of individual imaging parameters for classification, we ranked MRI features according to their contribution weights across the 20 runs. Higher contribution weights indicate a greater importance of features in distinguishing schizophrenia patients from healthy controls.

Antipsychotic Drug Effects and Illness Duration Effects on Classification

We compared classification accuracy across the 4 patient-control pairs of interest using χ² tests, including comparisons of untreated and treated groups (separately for early and later course of illness patients, ie, NT-FES vs AT-FES, and NT-LTSZ vs AT-LTSZ) and early and later course of illness groups (separately for untreated and treated groups, i.e., NT-FES vs NT-LTSZ, and AT-FES vs AT-LTSZ) to examine the effects of antipsychotic medication and illness duration respectively.

Results

Feature Extraction and Feature Selection

Frontal, temporal, and occipital cortical regions as well as the hippocampus were selected as relevant features in all 4 classifiers. Contribution of GMV in left hippocampus was observed regardless of treatment status and illness duration. Features of GMV in right superior temporal gyrus were useful for classification in treated FES patients and long-term ill patients, but not for the NT-FES group. Features useful for the 4 classification models are presented in Supplementary figure S1.

Classification Performance Across Patient Subgroups

The MKL models achieved classification accuracy for never-treated patients of 0.73 for NT-FES (p < .001, sensitivity = 0.73, specificity = 0.74) and 0.83 for NT-LTSZ (p < .001, sensitivity = 0.84, specificity = 0.82), and 0.94 for AT-FES (p < .001, sensitivity = 0.96, specificity = 0.93) and 0.98 for AT-LTSZ (p < .001, sensitivity = 0.97, specificity = 0.98) among antipsychotic-treated patients (figure 1). Classification performance was significantly different among the four patient-control classification pairs (χ² = 59.02, p < 0.001).

For case-control classification, the contribution of functional measures ranked first in the NT-FES vs healthy control pair, while structural measures rose to dominate in patients after antipsychotic treatment and with long-term illness duration. Of note, the most distinguishing features of untreated first-episode patients were functional physiological rather than anatomic features, suggesting that early in the illness course during acute illness episodes, pronounced functional alterations are the most prominent brain feature associated with illness, and anatomic features contribute more to individual case identification after treatment and with longer illness course.

Antipsychotic Drug Effects on Classification

In individual patient identification, treated patients both at early and later course of illness (AT-FES and AT-LTSZ) showed significantly higher classification accuracy compared to never-treated patients with similar illness duration (χ² = 20.54 and χ² = 17.81 respectively, both p < .001). These findings suggest that both early and later in the illness course, antipsychotic treatment has impacts on brain features that increase the separation of individuals with schizophrenia from HCs beyond the differences resulting from direct illness effects.

Next, we aimed to identify features that contribute to patient classification before and after treatment. Treated patients with both short- and long-term illness duration showed feature contributions of SA of left lateral orbitofrontal cortex, CT of left pars triangularis, right medial orbitofrontal cortex and right pars opercularis, GMV of right putamen, left inferior and medial frontal gyrus, bilateral middle temporal gyrus and left amygdala, and fALFF of right superior frontal gyrus (figure 2). None of these features were included as key relevant features in either untreated group. Selected features for each of the four case-control pairs are presented in Supplementary figure S1.

Illness Duration Effects on Classification

With regards to illness course effects, classification accuracy was marginally higher in case-control differentiation for treated patients with long-term illness course relative to first-episode patients who had received 1-year antipsychotic treatment (AT-LTSZ vs. AT-FES, χ² = 4.55, p = .03). Classification performance did not significantly differ between never-treated long-term ill and first-episode patients (NT-LTSZ vs NT-FES, χ² = 2.70, p = 0.10).

Long-term ill patients irrespective of treatment were characterized by alterations of SA of left inferior temporal cortex, CT of right pars triangularis region, GMV of right hippocampus, bilateral calcarine cortex and left middle occipital regions, fALFF of left inferior occipital cortex, and ReHo of left rolandic operculum and left angular gyrus. None of these features were selected in short-term ill patients regardless of treatment status (figure 2).

Discussion

By including both untreated and treated patients with either short- or long-term illness duration who were all scanned on the same scanner with a uniform data acquisition and analysis protocol, the present MKL study using both anatomic and functional brain features examined treatment status and illness duration effects on classification performance discriminating schizophrenia patients from healthy controls. This approach provided a simultaneous consideration of all brain features to maximally separate groups and identify brain features relevant for case-control discrimination. Our results demonstrated the robust contribution of treatment status and modest contribution of illness duration on differentiating schizophrenia patients from healthy controls. Thus, previously reported effects of antipsychotic treatments on brain anatomy¹⁰ and function⁹ appear to alter brain features in ways that significantly improve case-control discrimination. This effect is seen most convincingly in the direct comparison of diagnostic discrimination of first-episode patients before treatment and after 1-year treatment.

The observation that multimodal MRI features provide only moderate utility for discriminating untreated schizophrenia patients from healthy individuals implies that intrinsic illness-related brain alterations are not at a level of consistency and severity suitable to enable reliable and accurate identification of schizophrenia patients for clinical purposes. These effects were demonstrated both in never-treated first-episode patients and in long-term ill patients even when both functional and anatomic features are considered simultaneously. The more successful identification of treated schizophrenia patients relative to controls, both early and later in illness course, appears to result from leveraging the impact of antipsychotic drugs on brain anatomic and functional features to increase case identification accuracy.

Antipsychotic medications treatment, commonly employed as a therapy for many psychiatric disorders in addition to schizophrenia, such as bipolar disorder, major depressive disorder, and autism, appears to have diverse effects on regional brain anatomy and function of sufficient degree to assist quite meaningfully in the consistent identification of individual schizophrenia patients. Drugs in this class can have complex pharmacology, but all are dopamine receptor (D2) antagonists,²⁷ and thus act preferentially in brain regions that most robustly express dopamine receptors.^28,29 Potentially for this reason, the features that were most useful for classification purposes in treated patients included anatomic alterations in putamen, amygdala, and frontal cortex that were seen after 1-year treatment in the first-episode cohort compared to their pretreatment baseline. A similar pattern of effects was seen in treated long-term ill patients relative to untreated long-term ill patients. These findings suggest that MRI features may be useful for studying antipsychotic effects on the brain, for which their potential use for comparing effects on the brain of different treatments and predicting clinical outcomes remains to be systematically investigated.^30,31 However, applying MRI features to study illness effects was limited by the use of antipsychotic medications, thus using first-episode drug-naïve schizophrenia patients can avoid treatment-related effects.³¹

Challenges for developing diagnostic imaging markers have been noted conceptually by an American Psychiatric Association work group.³² Our findings identify 2 concerns in that regard, first the limited utility in untreated schizophrenia patients even in case-control comparisons, and second the demonstrations that drug effects and, to a lesser degree, course of illness effects significantly affect biomarker performance. In this context, alternative approaches to biomarker development need to be considered. Because schizophrenia is a complex and variable behaviorally defined illness, MRI biomarkers may prove more useful for delineating illness subtypes than for general illness identification.^33–35 Second, more novel MRI features may be needed to provide more useful biomarkers, such as different acquisition approaches such as diffusion tensor imaging (DTI) and magnetic resonance spectroscopy (MRS), and novel image processing procedures such as new texture analysis approaches.³⁶

Previous studies have shown progressive brain alterations associated with illness course in schizophrenia.^14,15,37 Our data showed similar modest effects, with discrimination of case and control participants being modestly greater in treated patients with longer-term than recent onset illness. However, this effect was not seen in the comparison of untreated patients in relation to illness duration. Thus, in contrast to robust medication effects on brain features, performance was less robustly influenced by illness progression effects. We note that modeling is complicated by the smaller size of the sample of individuals with untreated long-term illnesses. Yet, we note that case-control discrimination was not significantly different in never-treated patients close to illness onset and those scanned years after illness onset.

While our study identified brain measurements that enhanced classification performance after antipsychotic therapy, GMV of the hippocampus maintained its utility in classification regardless of illness duration and treatment history, including the comparison of FES patients before and after treatment. This is in accordance with previous studies showing hippocampus volume alterations in drug-naïve first-episode schizophrenia³⁸ and in long-term schizophrenia,³⁹ as well as accelerated hippocampal subfield volume loss in never-treated schizophrenia.⁴⁰ These findings are consistent with the view that anatomical deficits in the hippocampus are core features in the neuropathology in schizophrenia.^41,42

Our multimodal analysis provided the opportunity to contrast the discriminatory power of different anatomic and functional MR features, which is an advance in this field that has typically focused on a single modality or image feature. We found that functional measurements were the leading discriminating feature in NT-FES patient identification, while the contribution role for structural measurements rose with illness progression and in antipsychotic-treated individuals. The greater relevance of physiological measures for case-control discrimination at illness onset is consistent with previous studies showing robust functional deficits related to acute psychosis in first-episode schizophrenia that is reduced by treatment and clinical stabilization (albeit with some apparently adverse effects on activity in frontal systems).^9,43 In contrast, relatively long-lasting structural abnormalities that can increase somewhat over the long-term course of schizophrenia^15,44 are more common features later in the illness course. Thus, multimodal imaging may provide complementary and comprehensive information in schizophrenia patients with varying clinical relevance depending on the illness duration and treatment state of the patient.³²

Several limitations need to be considered when interpreting our results. First, the classification accuracy differentiating NT-FES patients from healthy controls (0.73) is lower than that of previously published studies,⁴⁵ which may be attributed to the heterogeneity of this disorder early in its course, or different sample sizes and imaging modalities used for case identification. Second, the sample sizes used in our four classification analyses were not the same, so the robustness of classification performance and feature identification may vary across the 4 case-control pairwise comparisons. This consideration is particularly relevant regarding the NT-LTSZ sample, which though a rare population and especially useful in the current study, was smaller than the other groups. Finally, this study did not recruit independent samples for validation. Our findings need to be replicated in independent samples.

This study provides important insights regarding the effects of antipsychotic medication and illness duration on brain anatomy and function in individual schizophrenia patients. Our results demonstrate a robust contribution of antipsychotic treatment status and a modest contribution of illness duration on the utility of MR features. The limited case-control discrimination in never-treated patients considering both anatomic and functional features, and the observation that robust classification is only achieved when effects of antipsychotic drugs on brain features are used to supplement direct effects of illness, are consistent with the view that psychiatric imaging, while a valuable tool for illness investigation, does not currently demonstrate standards of clinical utility as a diagnostic aid. Our findings demonstrating the robust effects of antipsychotic treatment on brain features raise issues beyond concerns about using ML approaches with MRI data to identify treated schizophrenia patients, to broader questions for MRI studies aiming to characterize illness-related brain pathology in treated patient samples unless parallel findings are observed in never-treated patients, and potentially about similar questions about treatment effects for studies of other psychiatric disorders.

Acknowledgments

We thank for all psychiatry staff responsible for patient recruitment and data collection. Dr. Sweeney acknowledges support from the University of Cincinnati Schizophrenia Program Fund. Dr. Lui acknowledges the support from Chang Jiang Scholars (Program No. T2019069) and support from Humboldt Foundation Friedrich Wilhelm Bessel Research Award.

Conflicts of Interest

Drs. Zhang, Li (S) and Sweeney consult to VeraSci. The remaining authors have declared that they have no conflicts of interest in relation to the subject of this study.

Funding

This study was supported by the National Natural Science Foundation of China (Project Nos. 82120108014, 82071908, and 81621003 to SL, and 82101998 to WZ), Chinese Academy of Medical Sciences (Project No. 2021-I2M-C&T-A-022 to SL), Sichuan Science and Technology Program (Grant Nos. 2021JDTD0002 to SL and 2020YJ0018 to WZ), the Science and Technology Project of the Health Planning Committee of Sichuan (Grant No. 20PJ010 to WZ), Post-Doctor Research Project, West China Hospital, Sichuan University (Grant No. 2020HXBH005 to WZ) and 1.3.5 Project for Disciplines of Excellence, West China Hospital, Sichuan University (Project Nos. ZYYC08001 and ZYJC18020 to SL).

References

Author notes