Eye Movement Characteristics for Predicting a Transition to Psychosis: Longitudinal Changes and Implications

Abstract

Background and hypothesis

Substantive inquiry into the predictive power of eye movement (EM) features for clinical high-risk (CHR) conversion and their longitudinal trajectories is currently sparse. This study aimed to investigate the efficiency of machine learning predictive models relying on EM indices and examine the longitudinal alterations of these indices across the temporal continuum.

Study design

EM assessments (fixation stability, free-viewing, and smooth pursuit tasks) were performed on 140 CHR and 98 healthy control participants at baseline, followed by a 1-year longitudinal observational study. We adopted Cox regression analysis and constructed random forest prediction models. We also employed linear mixed-effects models (LMMs) to analyze longitudinal changes of indices while stratifying by group and time.

Study results

Of the 123 CHR participants who underwent a 1-year clinical follow-up, 25 progressed to full-blown psychosis, while 98 remained non-converters. Compared with the non-converters, the converters exhibited prolonged fixation durations, decreased saccade amplitudes during the free-viewing task; larger saccades, and reduced velocity gain during the smooth pursuit task. Furthermore, based on 4 baseline EM measures, a random forest model classified converters and non-converters with an accuracy of 0.776 (95% CI: 0.633, 0.882). Finally, LMMs demonstrated no significant longitudinal alterations in the aforementioned indices among converters after 1 year.

Conclusions

Aberrant EMs may precede psychosis onset and remain stable after 1 year, and applying eye-tracking technology combined with a modeling approach could potentially aid in predicting CHRs evolution into overt psychosis.

Introduction

In early intervention studies of psychosis, the clinical high risk for psychosis syndrome (CHR)^1,2 is now used to describe the potential risk stage for the development of psychotic disorders, and individuals in this stage exhibit attenuated psychotic symptoms and reduced functioning, but do not yet meet the criteria for first-episode psychosis.³ The results of the Chinese epidemiological survey showed that the detection rate of CHR was 4.2% among outpatients who first sought help⁴ and about 1.4% among general college students.⁵ Fourteen to 21 years of age was the peak age group and the proportion of minors was 50%.⁴ Due to the poor understanding of the pathological mechanisms of CHR, the highly variable clinical symptoms, and individual heterogeneity in the CHR population,^6,7 psychopathology assessed by symptom rating instruments remains the only diagnostic strategy available. Recent studies have adopted multivariate prediction models or risk calculators, which may remain limited by the lack of specificity, principal reliance on clinical or other primarily phenomenological data, and the difficulty of rapid application due to the substantial standardized training and specialization.⁸ A review⁹ demonstrated that combining or sequentially using clinical data with neurobiological and physiological indexes optimizes risk prediction among CHR populations. Therefore, combining biomarkers with multivariate prediction models and risk calculators can help address the limitations mentioned above.⁸

Recent research has indicated that eye movement (EM) abnormalities, an endophenotype of schizophrenia,¹⁰ are also detected in CHR populations. Initial studies focused on the performances of CHR under saccadic tasks and found that the error rates of anti-,³ pro-,¹¹ and memory-guided saccades,¹² were higher than that of healthy controls (HC). However, 2 studies compared the anti-saccade accuracies of the CHR and HC groups but reported negative results.^12,13 Saccade latency as another critical indicator has been discovered to be prolonged in CHR individuals.¹⁴ In addition, EM deficits of CHR also appeared in other paradigms, which were specifically manifested as more and larger saccades,¹⁵ more and shorter fixations¹⁶ under the fixation stability task; increased corrective and non-corrective saccadic rates,¹⁷ as well as lower velocity gains¹⁶ under the smooth pursuit test; and fewer fixations and saccades, prolonged fixation duration and shortened saccade amplitude, along with decreased scan path lengths and saccade velocity under the free-viewing task.^15,16,18

Four studies have explored the ability of EM measures to diagnose and predict clinical conversion. Caldani et al. reported a receiver operating characteristic (ROC) area of 0.77 to distinguish 23 CHR individuals from 47 HCs by memory-guided error rates.¹² In addition, a recent study reported that the gradient-boosted decision tree model developed by Lyu et al.¹⁶ identified 88% CHR cases (16 CHR participants vs 104 HCs). Furthermore, Kleineidam et al.¹³ found that the indexes obtained from the anti- and pro-saccade tasks failed to predict clinical outcomes after 2 years (23 converters vs 137 non-converters). However, a recent study of ours,¹⁵ which included 21 converters and 87 non-converters, discovered that the saccade frequency from the fixation stability task and the saccade amplitude from the free-viewing test could predict the psychosis transition within 3 years, with an area under the logistic regression ROC curve of 0.80. Therefore, the performance of machine learning prediction models based on multidimensional EM indexes along with the longitudinal trajectory of EM characteristics of CHR populations remain unclear and necessitate further in-depth inquiry.

Building on existing research,¹⁵ we aimed to investigate EM characteristics of CHR under 3 tasks at baseline and after 1 year of follow-up, explore EM measurements associated with CHR conversion, establish machine learning prediction models, and evaluate the longitudinal trajectory of EM metrics.

Methods

Participants

This study was approved by the Ethics Committee of Shanghai Mental Health Center. All participants signed the informed consent (minors were required to sign both by themselves and their guardians).

One hundred and forty CHR individuals were enrolled in the outpatient department of Shanghai Mental Health Center. The inclusion criteria: (a) aged between 15 and 45 years old; (b) more than 6 years of education; (c) compliance with the CHR criteria was evaluated by a professional clinician using the Structured Interview for Psychosis-risk Syndromes (SIPS) and Scale of Psychosis-risk Symptoms (SOPS) assessment. Participants were diagnosed with at least 1 of 3 prodromal syndromes, including brief intermittent psychotic syndrome, attenuated positive symptom syndrome, or genetic risk and deterioration syndrome; (d) drug-naïve before entering the study. Exclusion criteria encompassed the following: (a) with DSM-IV-I axis psychiatric disorders, such as schizophrenia, affective disorders, and anxiety spectrum disorders, assessed using the Mini International Neuropsychiatric Interview (MINI); (b) IQ score below 70, as measured by the Wechsler Intelligence Scale for Children/Adult-Chinese Revised^19,20 during the enrollment screening stage; (c) with severe medical diseases, such as neurological disorders, cancer, heart failure, immune deficiency; (d) with symptoms caused by drug abuse or dependence; (e) with audio-visual disorders (abnormal hearing, high myopia, color blindness, and other eye diseases that affected vision).

Ninety-eight HCs were recruited through online advertising in Shanghai and screened to exclude the presence of any mental disorders or prodromal symptoms/syndromes using MINI and SIPS. The inclusion and exclusion criteria for HC were the same as for CHR, except that HCs did not meet the criteria for the prodrome syndrome.

Conversion to Psychosis

At follow-up, based on information from face-to-face or telephone interviews (patient and family reports) along with clinician reports, the clinical outcome of CHR was gauged via the operational definition of SIPS psychotic episodes. Specifically, the conversion was defined as the presence of at least 1 psychotic symptom (a score of “6” on at least 1 item on the positive symptom scale) at the 1-year follow-up time point.²¹

EM Measures

Fixation Stability Test

The participants were asked to keep looking at a black dot (size 0.5°) in the center of the computer screen and ignore interference signals (“*” mark, size 0.5°) from the side. During the first 5 s, only the black dot appeared (simple task); for 5–10 s, the distractor appeared randomly (direction: left and right; distance: near and far) next to the black dot for a specified duration (distracted task). There were 10 trials and 4 indexes were analyzed, including fixation number, fixation duration, saccade number, and saccade amplitude.

Free-viewing Test

The participants were instructed to relax and freely view 35 still black and white pictures, with each presented for 10 s. The pictures involved 4 types of content: natural scenery, social scenes, still objects, and meaningless images, and were presented in a random order. We analyzed 6 variables: the fixation number, the fixation duration, the saccade number, the saccade amplitude, the scan path length, and the saccade velocity.

Smooth Pursuit Test

The participants were required to keep their heads motionless and their eyes to follow the black dots on the screen, which moved at a certain frequency and trajectory. The target moved horizontally and vertically with a sinusoidal trajectory for 2.5 s each (horizontal/vertical pursuit paradigm), or it moved both horizontally and vertically with Lissajous trajectories of different frequencies (slow Lissajous paradigm: 20 s; fast Lissajous paradigm: 10 s). Each trial was repeated once. We measured the horizontal velocity gain, the vertical velocity gain, the fixation number, the saccade number, and the saccade amplitude. See supplementary materials, table S1 for the explanations for all EM variables.

EM Tracking

Eye Link 1000 desktop eye tracker from SR Research Canada was applied to collect EM data. The eye tracker set to pupillary reflex tracking mode had a 17-in. display with a resolution of 1024 × 768 pixels. An adjustable chair was positioned 90 cm in front of the display, and a jaw rest stabilized the participant’s head to align their eyes with the center of the display. EM data were tracked within a range of 30± degrees horizontally and vertically. Furthermore, data were collected from dominant eyes with a 500 Hz sampling frequency.

Data Analysis

Prior to the analysis, EM data cleaning was conducted by eliminating blink artifacts, fixations with duration below 20 ms or outside the screen, and saccades moving from inside or outside the screen.²² Univariate extreme outliers were identified by evaluating histograms and box plots, and the Winsorization method was used to replace the outliers with the nearest highest or lowest observation to minimize the impact on the results of the parametric statistical tests.²³

Statistical analysis was performed using R Studio software, version 4.3.1. Demographic and clinical data were compared using a 1-way ANOVA/independent t-test for continuous variables and a chi-square test for categorical variables. One-way Multivariate Analysis of Variance (MANOVA) was used to compare differences in EM variables between groups, followed by post hoc multiple comparisons using the Bonferroni correction (the effect size being measured by Cohen d). Cox regression with forward stepwise selection (likelihood ratio) was applied for survival analysis.

The caret and e1071 packages of R were applied for feature selection and machine learning model construction. All EM variables were used as features and underwent column normalization as a preprocessing step before learning to improve convergence speed. The recursive feature elimination (RFE) algorithm coupled with random forest (RF) (5-fold cross-validation) was employed as the feature selection function to identify the most predictive subset of features. Next, the new data set was randomly divided into a training data set with 60% of the sample and a test data set with 40% of the sample.

The training data were balanced with the Synthetic Minority Oversampling Technique (SMOTE) via the DMwR package of R. SMOTE is a data augmentation technique that creates synthetic samples based on information from existing data, effectively addressing class imbalance and promoting balanced representation across categories with varying observations. This method helps achieve a more robust and reliable predictive model by mitigating the impact of imbalanced datasets.²⁴ The perc. over parameter was set to 200%, resulting in the positive class samples being oversampled by 200%. Additionally, the k parameter was set to 5, implying that the latest 5 samples of each positive category were selected for replication.

We employed RF classification models to distinguish between CHR-C and CHR-NC, and the optimal parameter combination was determined through grid search and cross-validation techniques. Accuracy, sensitivity, specificity, and the area under the ROC curve were adopted to evaluate the performance of the model on the test data set.

Linear mixed-effects models (LMMs) were conducted to examine the different groups’ longitudinal trajectories in EM measures via R package lme4. When fitting the LMM, we considered group (CHR-C, CHR-NC, and HC), time (baseline, follow-up), and their interaction as fixed effects, and subject ID as a random effect. Additionally, baseline gender, age, education year, along with dosage, and duration of antipsychotic use within the past year were included as covariates for control. The simple effects were analyzed by R package emmeans with P values adjusted using Bonferroni correction.

Lastly, partial correlation analysis was used to reveal the relationship between baseline EM characteristics and clinical symptoms, with age, gender, and education years controlled and the false discovery rate corrected.

Result

Sample Characteristics

At baseline, 140 drug-naïve CHR participants and 98 HCs were enrolled. 123 CHR participants (87.86%) completed a 1-year clinical follow-up. Among them, 25 cases were converted to psychosis (CHR-C) (15 were hospitalized for treatment during a first psychotic episode, 60%), and 98 cases were not converted (CHR-NC), with a conversion rate of 20.33%. There were no significant differences in demographic data or clinical symptoms between CHR cases who completed follow-up (n = 123) and those who were lost to follow-up (n = 17).

The CHR-C, CHR-NC, and HC groups were matched in age (F = 2.293, P = .103), gender (χ² = 1.000, P = .606), and years of education (F = 2.831, P = .061) at baseline. Independent sample T-test demonstrated that the CHR-C group exhibited significantly higher scores for negative symptoms (t = 3.903, P < .001), disorganized symptoms (t = 2.234, P = .027), and total SOPS score (t = 2.737, P = .007) compared with the CHR-NC group.

At the follow-up assessment, 21 CHR-Cs (85%) and 57 CHR-NCs (58.16%) reported the use of antipsychotics within 1 year. Notably, there were no significant group differences in daily dosage (equivalent use of olanzapine²⁵) (t = −1.105, P = .273), but distinctions did emerge in the duration of usage (t = 2.608, P = .011) (see more information on medication in supplementary materials). Demographic data and clinical characteristics of the 3 groups are shown in table 1.

Table 1.

Demographic and Clinical Characteristics of the 3 Groups

	CHR-C (n = 25)	CHR-NC (n = 98)	HC (n = 98)	F/t/χ 2	P value
Baseline demographic and clinical characteristics
Age, years, M (SD)	18.44 (5.27)	20.21 (5.92)	18.96 (3.19)	2.293	.103
Gender (male/female)	15/10	48/50	49/49	1.000	.606
Education year, years, M (SD)	10.56 (2.50)	11.92 (3.24)	12.14 (2.82)	2.831	.061
SOPS positive scores, M (SD)	9.20 (3.35)	9.28 (4.19)	—	−0.083	.934
SOPS negative scores, M (SD)	15.56 (6.89)	10.36 (5.69)	—	3.903	<.001***
SOPS disorganized scores, M (SD)	7.12 (3.33)	5.44 (3.36)	—	2.234	.027*
SOPS general scores, M (SD)	9.40 (3.03)	8.48 (3.37)	—	1.244	.216
SOPS total scores, M (SD)	41.28 (10.73)	33.55 (13.03)	—	2.737	.007**
Follow-up medication information
Use of antipsychotics, n (%)a	21 (85%)	57 (58.16%)	—	5.731	.017*
Daily antipsychotic dose in OLAequ, mg/d, M (SD)	12.59 (9.82)	10.16 (8.13)		−1.105	.273
Use duration of antipsychotics within 1 year, weeks, M (SD)	29.24 (18.98)	39.72 (14.41)	—	2.608	.011*

^aMore information on follow-up medication can be found in the supplementary material.

Abbreviations: CHR, clinical high risk; CHR-C, CHR-converter CHR-NC, CHR-nonconverter; HC, healthy control; SOPS, Scale of Psychosis-risk Symptoms; OLAequ, equivalent use of olanzapine. *P < .05, **P < .01, ***P < .001.

EM Characteristics

MANOVA reported that 13 indexes differed by the group, including 2 of the fixation stability test, 5 of the free-viewing test, and 6 of the smooth pursuit test (table 2). See supplementary materials, table S2 for all EM variable information. Figure 1 gives representative examples of the differences in fixation trajectories and distribution among the 3 groups.

Table 2.

Eye Movement Variables with Significant Between-group Effects

	CHR-C (n = 25)	CHR-NC (n = 98)	HC (n = 98)	F value	P value	Post hoc contrast Cohen d (Bonferroni corrected P value)
						CHR-C vs CHR-NC	CHR-C vs HC	CHR-NC vs HC
Fixation stability (simple)
Saccade amplitude	1.78 (1.02)	2.16 (1.70)	1.45 (1.11)	6.394	.002**	0.27 (.680)	0.31 (.862)	0.49 (.001**)
Fixation stability (distracted)
Saccade amplitude	1.89 (0.67)	2.15 (1.26)	1.61 (0.84)	6.570	.002**	0.26 (.811)	0.37 (.688)	0.50 (.001**)
Free-viewing test
Fixation number	20.32 (4.91)	22.15 (4.30)	22.48 (3.16)	3.059	.049*	0.40 (.114)	0.52 (.043*)	0.09 (1.000)
Fixation duration	354.74 (99.37)	307.61 (60.85)	313.13 (71.79)	4.489	.012*	0.57 (.010*)	0.48 (.028*)	0.08 (1.000)
Saccade number	20.55 (4.63)	22.80 (4.97)	23.63 (3.94)	4.708	.010*	0.47 (.079)	0.72 (.008**)	0.19 (.605)
Saccade amplitude	4.51 (1.26)	5.38 (0.92)	5.15 (0.90)	8.244	<.001***	0.79 (<.001***)	0.58 (.010*)	0.25 (.290)
Scan path length	103.00 (37.09)	120.39 (33.55)	122.27 (30.80)	3.546	.031*	0.49 (.056)	0.57 (.028*)	0.06 (1.000)
Smooth pursuit test (vertical)
Vertical velocity gain	0.40 (0.13)	0.48 (0.16)	0.49 (0.16)	3.227	.042*	0.55 (.087)	0.62 (.037*)	0.06 (1.000)
Saccade number	8.64 (1.82)	8.34 (1.95)	7.78 (1.79)	3.303	.039*	0.16 (1.000)	0.48 (.119)	0.30 (.109)
Smooth pursuit test (horizontal)
Saccade amplitude	2.98 (1.36)	2.39 (0.92)	2.58 (1.10)	3.238	.041*	0.51 (.039*)	0.32 (.267)	0.19 (.645)
Smooth pursuit test (slow Lissajous)
Horizontal velocity gain	0.89 (0.09)	0.91 (0.09)	0.93 (0.08)	3.286	.039*	0.22 (.605)	0.47 (.057)	0.23 (.274)
Vertical velocity gain	0.82 (0.09)	0.88 (0.09)	0.88 (0.11)	3.913	.021*	0.67 (.033*)	0.60 (.021*)	0 (1.000)
Smooth pursuit test (fast Lissajous)
Vertical velocity gain	0.67 (0.15)	0.74 (0.14)	0.75 (0.14)	3.172	.044*	0.48 (.090)	0.55 (.040*)	0.07 (1.000)

Note: This table only shows the eye movement variables that differed significantly between the 3 groups, and see supplementary table S2 for other variable information. Continuous variables are expressed as M (SD).

Abbreviations: CHR, clinical high risk; CHR-NC, CHR-nonconverter, CHR-C, CHR-converter, HC, healthy controls. Bolded numbers indicate statistical significance: *P < .05, **P < .01, ***P < .001.

Fig. 1.

Illustrative fixation stability (distracted task), free-viewing, and smooth pursuit (slow Lissajous). These example trials illustrate the types of idiosyncrasies, features, and errors that are commonly observed amongst the respective groups of participants.

The post hoc analysis showed that 11 indexes survived after the Bonferroni correction. For the fixation stability test, compared with HCs, the CHR-NC group showed higher saccade amplitudes in both simple and distracted conditions (Cohen d = 0.49, P = .001; Cohen d = 0.50, P = .001). Follow-up comparison of variables of the free-viewing test revealed fewer saccades and fixation, decreased saccade amplitude and scan path length, and longer fixation duration of the CHR-C group relative to HCs (Cohen d = 0.48–0.72, P < .05). Furthermore, the CHR-C group showed prolonged fixation duration and reduced saccade amplitude (Cohen d = 0.57, P = .010; Cohen d = 0.79, P < .001).

As for the smooth pursuit task, the CHR-C group manifested lower vertical velocity gain than HCs under the vertical, slow, and fast Lissajous pursuit tasks (Cohen d = 0.55–0.62, P < .05). In addition, increased saccade amplitude (horizontal) and lower vertical velocity gain (slow Lissajous) were detected in the CHR-C group compared with the CHR-NC group (Cohen d = 0.51, P = .039; Cohen d = 0.67, P = .033).

Prediction of EM Measures for Psychosis Conversion

Cox regression with forward stepwise selection (likelihood ratio) was used to explore the EM variables associated with CHR conversion. Four EM measures differed between the CHR-C and CHR-NC groups were chosen as possible candidates, including the fixation duration and the saccade amplitude from the free-viewing test, the saccade amplitude from the Horizontal pursuit task, and the vertical velocity gain from the slow Lissajous pursuit task. The analysis reduced the 4 measures to 3, which were the fixation duration (β = 0.008, OR = 1.008, P = .004) and the saccade amplitude (β = −0.707, OR = 0.493, P = .001) from the free-viewing test, as well as the vertical velocity gain (β = −4.313, OR = 0.010, P = .018) from the slow Lissajous pursuit (supplementary materials, table S3).

Prediction Models

RFE combined with RF revealed that the model based on the 4 predictive features could achieve accurate predictions (supplementary materials, figure S1). Subsequently, we selected a specific subset of the 4 features based on their F-scores, namely the fixation duration and the saccade amplitude (the free-viewing test), the saccade amplitude (horizontal pursuit), and the vertical velocity gain (slow Lissajous pursuit), to construct a new dataset for subsequent modeling and analysis.

Upon performing grid search and cross-validation, the RF model’s optimal parameters were identified as follows: Num Trees: 300, Max Depth: 9, and Max Num Splits: 1. The performance of the model on the test data set is as follows: an accuracy of 0.776 (95% CI: 0.633, 0.882), a sensitivity of 0.800, a specificity of 0.769, and a ROC area of 0.854.

In addition, we also evaluated the predictive performance of EM variables combined with demographic and clinical data. The selected features included the above 4 EM variables, age, gender, education year, SOPS total, and subscale scores. Similarly, the data set was divided into a training data set (60%) and a test data set (40%). The hyperparameters were optimized with the following values: Num Trees: 300, Max Depth: 5, and Max Num Splits: 1. The RF model achieved an accuracy of 0.776 (95% CI: 0.634, 0.882), a sensitivity of 0.700, a specificity of 0.795, and a ROC area of 0.854 on the test set. Figure 2 displays the ROC curves of 2 classification models.

Fig. 2.

ROC curves of the prediction models. Eye movement data refers to a combination of 4 eye movement features. Eye movement, demographic, and clinical data refers to the 4 eye movement features, age, gender, education year, SOPS total score, and subscale scores.

Longitudinal Changes in EM Variables

A total of 114 participants (CHR-C: n = 12; CHR-NC: n = 67; HC: n = 35) underwent a 1-year follow-up for EM assessments, with a completion rate of 51.58% (114/221). Results from the linear mixed models revealed significant main effects of time in 3 EM variables, including the fixation number (β = 1.06, SE = 0.51, P = .040) and the saccade number (β = 1.05, SE = 0.53, P = .049) of the distracted fixation stability test, as well as the scan path length (β = −13.82, SE = 4.60, P = .003) from the free-viewing test.

Furthermore, there were significant group × time interaction effects from 1 year to baseline between the CHR-NC and HC groups in 3 EM measures, including the saccade amplitude (β = −0.71, SE = 0.26, P = .007) from the simple fixation stability test, the saccade amplitude (β = −0.64, SE = 0.19, P < .001) and the scan path length (β = −12.09, SE = 5.74, P = .037) from the free-viewing test. See supplementary materials, table S4 for the linear mixed effects of all EM measures.

The simple effects pairwise comparisons revealed a significant decrease in saccade amplitudes (simple fixation stability; free-viewing) within the CHR-NC group after 1 year from baseline (baseline—1 year: β = 0.64, SE = 0.16, P < .001; β = 0.70, SE = 0.12, P < .001, Bonferroni corrected). The scan path length of all 3 groups showed a significant reduction after 1 year (CHR-C: β = 21.30, SE = 8.08, P = .009; CHR-NC: β = 25.90, SE = 3.53, P < .001; HC: β = 13.80, SE = 4.67, P = .004, Bonferroni corrected) (figure 3).

Fig. 3.

Estimated marginal means of variables with significant interactions. Error bars indicate the standard error. Abbreviations: CHR, clinical high risk; CHR-C, CHR-Converters; CHR-NC, CHR-Non converters; FS(S), the fixation stability test (simple); FV, free-viewing test; HC, healthy controls.

Relationship Between Baseline EM Characteristics and Clinical Symptoms

With age, gender, and education years controlled and the false discovery rate corrected, only the general symptom score of the CHR-NC group was negatively correlated with the saccade velocity of the free-viewing test (r = −0.375, P < .001) (supplementary materials, tables S5–7).

Discussion

The present study demonstrated that free-viewing and smooth pursuit abnormalities were associated with the clinical progression of CHR, accentuating their potential utility as promising biomarkers. Furthermore, the performance of the EM-based prediction models was relatively outstanding, and the EM alterations of the CHR-C group remained relatively stable over the course of 1 year.

The role of fixation is to fix the image of the target stimulus in the fovea,²⁶ examining the individual’s attentional control. Prior studies reported more and larger saccades in CHR,^15,16 but in this study we observed an increase in saccade amplitude specifically within the CHR-NC group, implying that saccade abnormalities may not yet be evident in CHR-C at baseline. Furthermore, maintenance of fixation stability involved visual areas (V2 and V4)²⁷ and visual processing and sensorimotor transformation,²⁸ and abnormalities in these structures and functions have been reported in schizophrenia.^29,30 Furthermore, maintaining stable fixations might mirror the visual field dependence ability and involve the visuospatial cognitive function, which was found to be impaired in CHR.³¹

The free-viewing task embodied visual search features³² and involved more complex cognitive processes than the other 2 paradigms,³³ and the free-viewing mode was divided into the focal and the ambient modes.³⁴ Our study found that the CHR-C group might adopt the focal mode, as evidenced by fewer fixations and saccades, prolonged fixation duration, shorter saccade amplitude and scan path length, which was consistent with previous studies.^14–16 Of note, the fixation duration and the saccade amplitude also differed between the CHR-C and the CHR-NC groups. Similar to patients with schizophrenia,³⁵ those at CHR might depend more on focal patterns for local scanning than global exploration.¹⁸ Reliance on the focal mode might be attributed to the impaired top-down attention mechanisms,³⁶ and 1 study³⁷ exploring mechanisms of lateralization of language function and auditory attention impairment in CHR suggested that the top-down attention mechanism of CHR might be impaired. Moreover, the ambient mode involved the dorsal visual pathway that was impaired in schizophrenia,³⁸ which might also account for CHR’s reliance on focal scanning.

The smooth pursuit predicted the speed and direction of the target and adjusted the speed and amplitude of EMs to ensure that the target was positioned in the fovea central region of the retina.³³ Partially aligning with prior findings,^16,17 the CHR populations manifested increased saccade amplitudes and reduced velocity gain. Of note, the above alterations might vary depending on the trajectory of movement exhibited by the target stimulus. Reduced velocity gains meant a poorer ability to reconcile EM velocity and target speed,³³ and larger saccade amplitudes implied a deficiency in saccade suppression.¹⁷ Abnormalities in brain structures such as area V5 (involved in motion perception)³⁹ and frontal visual field (involved in target prediction)⁴⁰ were associated with smooth pursuit deficits in schizophrenia. Structural abnormalities in the frontal lobe were also found among the CHR populations, which might be related to the observed abnormal smooth pursuits.¹⁷

Cox regression analysis found 3 predictors significantly associated with higher risk of psychosis conversion, including the fixation duration and the saccade amplitude of the free-viewing task, and the vertical velocity gain from the slow Lissajous pursuit. The confined scan pattern of free-viewing represented a broad defect in schizophrenia and contributed the most to distinguishing schizophrenia from HCs.^41–43 This view corresponded to the current study: 2 predictors were derived from the free-viewing test.

Studies have highlighted the effectiveness of position and velocity gains obtained from the smooth pursuits in differentiating between schizophrenia patients and HCs. The calculation of EM variables is sensitive, so slight modifications to the paradigm location^22,44 and velocity gains⁴⁵ were effective in the diagnosis of schizophrenia. The sensitivity of EM variables meant that even minor changes to experimental conditions could significantly affect results,⁴⁶ underscoring the need for more refined indicators to accurately detect biomarkers.

The prediction models constructed using the EM parameters performed well, with an accuracy of 0.776 and a ROC area of 0.854. The feature screening analysis showed that the 4 EM measures served as crucial factors in the prediction of clinical transition. The model in this study outperformed our previous prediction models,¹⁵ possibly due to different metrics and larger sample size, while also noting that a more sensitive and objective machine learning approach⁴⁷ might help facilitate the discrimination.²² In short, the application of RFE for feature selection (note that these features differed between groups), coupled with dataset partitioning and cross-validation, oversampling techniques, and hyperparameter tuning collectively emphasized the robustness of the model in forecasting the evolution into psychosis.

As previous research has indicated,¹⁵ the model accuracy did not demonstrate a significant improvement upon the simultaneous inclusion of combined demographic, clinical, and EM variables. There are 3 explanations for the observed result: (a) EM-based models were relatively robust, (b) previous studies have shown that the demographic and clinical data were primarily predictive of outcomes 2–3 years later,⁴⁸ so their predictability might not be evident during the 1-year follow-up period, and (c) it is plausible that the limited inclusion of demographic and clinical data might not have been adequate to enhance the predictive performance.

One major contribution of this study is the comprehensive analysis of how EM characteristics evolve over time, revealing valuable insights into the underlying dynamics. Despite the lack of EM follow-up results for CHR, schizophrenia research essentially explained the observed results. In partial agreement with the schizophrenia patients,⁴⁹ the saccades of the simple fixation task of CHR-NC were temporally unstable, showing a decline after 1 year. Furthermore, the CHR-NC group also demonstrated a decrease in the saccade amplitude under the free-viewing task. Our Cox regression analysis showed that the saccade amplitude during the free-viewing task was a predictor in predicting conversion. As the study’s follow-up period was only 1 year, it is possible that this metric will be critical in identifying new CHR-C cases in future follow-ups.

In addition, the scan path length in all 3 groups showed a decline after 1 year, albeit with varying degrees, implying this metric showed variations across time, and this trend might reflect adjustments or adaptations in EM control and attention mechanisms within 1 year. The different degrees of change might be related to baseline EM characteristics, suggesting that this metric might play a significant role in disease progression. Notably, the smooth pursuit parameters displayed good temporal stability, thus strengthening the current view that this task could be used as endophenotypes.⁴⁹

Importantly, 3 factors should be accounted for the interpretation of the above longitudinal changes: the small sample size of CHR-C, the comparison of changes at only 2-time points which might not exclude natural fluctuations, and potential selection bias due to certain CHR cases not completing the EM follow-up.⁵⁰ Moreover, our findings demonstrated that the absence of significant longitudinal variations in other metrics for the CHR-C group suggested a relative stability in the EM impairments observed at baseline. In conclusion, EM impairments among CHR individuals were relatively stable over time, thereby bolstering the perspective that such aberrations were potential endophenotypic markers of schizophrenia.¹⁰

Several limitations need to be taken into account. Primarily, an incomplete follow-up assessment of EM might have introduced selection bias, reinforcing the need for larger sample sizes to validate future studies. Additionally, the predictive models in this investigation have not been externally validated, meriting future research to confirm their reproducibility across different data sets.

Conclusion

In summary, the aberrant manifestations of free-viewing and smooth pursuit were shown to be significantly correlated with the clinical transition among CHR individuals for psychosis, thereby highlighting their potential utility as accurate, presymptomatic biomarkers. Furthermore, it was observed that the EM anomalies within the CHR cohort exhibited sustained stability over the duration of 1 year. This study thus elucidates the plausibility and practicality of applying eye-tracking modalities for assessing the vulnerability of developing psychosis within a clinical context.

Supplementary Material

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

Funding

This work was supported by the National Natural Science Foundation of China (82151314, 81901832, 81871050, 82171497, and 82101623); Clinical Research Plan of SHDC (SHDC2022CRD026, SHDC2020CR4066).