We mainly collected ASD cases in Tianjin Anding Hospital from October 2018 to October 2023. The criteria of dropout (intolerance) mainly included: ①intolerance to the 1st rTMS intervention, such as adverse reactions to coil contact (distress or resistance) or intolerable side effects (including headache, dizziness, worsened anxiety, or seizures) leading to participant withdrawal or researcher-mandated discontinuation;② participants who have completed the first intervention but failed to the whole course (10 rTMS interventions). The eligibility criteria for dropout included: (1) it meets the diagnostic criteria of ASD in the fifth edition of the American Diagnostic and Statistical Manual of Mental Disorders (DSM-V); (2) age of 3–18 years old; (3) no medication during the rTMS intervention; (4) right-handed; (5) the total score of Childhood Autism Rating Scale (CARS) in the baseline ≥ 30 (22); and (6) less than 10 rTMS interventions. Similarly, ASD children who completed the whole rTMS intervention were selected as the control group with the same eligibility criteria except “complement at least two courses of rTMS intervention”, by matching chronological age, gender and socioeconomic status (SES). The exclusion criteria for both case and control group were: (1) contraindications to rTMS, such as metal or electronic instruments near the coil stimulation site; participants with a history of epilepsy (excluding epilepsy according to their electroencephalogram and medical record); participants with a history of brain trauma, brain tumors, and other diseases; participants with severe or recent heart disease; or other major physical illness. (2) Diagnosis of other mental illnesses (e.g., attention-deficit hyperactivity disorder, schizophrenia and depression). (3) Other neurodevelopmental disorders, genetic metabolic disease, or severe neurological disease. (4) Participants who could not cooperate with the eye movement experiment. (5)Participants who withdrew for non-medical reasons, such as scheduling conflicts or relocation, were not classified as intolerant. The study was conducted under the Code of Ethics of the World Medical Association (Declaration of Helsinki). Also, the study complied with all relevant national regulations and institutional policies and had been approved by the Medical Ethics Committee of Tianjin Medical University. Participants and their parents (or legal guardians) obtained all information about the research, including the purpose, requirements, responsibilities, compensation, risks, benefits, and alternatives. All questions were answered before asking for the consent signature.

We have used the self-made dropout questionnaire to investigate the parents of the dropout ASD patients through telephone or WeChat. The main contents included: demographic information of patients, clinical characteristics (such as the score of CARS, and IQ), reasons for withdrawal from treatment (Multiple-choice questions), the most intolerable thing for intervention (Single-choice Questions), and their feelings about rTMS intervention.

We evaluated the symptoms of ASD with CARS. The CARS consists of 14 domains assessing behaviors associated with autism, with a 15th domain “rating general impressions” of autism. Each field has a scale of one to four. Higher scores indicate a higher level of impairment. Total scores can range from 15 to 60. Scores below 30 mean that the individual is in the non-autistic range, a score between 30 and 36.5 indicates mild to moderate autism, and scores between 37 and 60 indicate severe autism.

The stimuli were selected from the Chinese Affective Picture System (23) (CAPS) and consisted of 48 different pictures (24 male faces and 24 female faces). Each picture included two black-and-white photographs of the same type portraying varied emotional valence (positive/negative + neutral) of the same person. Each picture was equal in size and symmetrical in position. There were three factors in this study, including the picture type (male, female), the left or right visual field where the emotional pictures presented (LVF, RVF), and the picture valence (positive, negative). There was one block for each condition and 6 trials in each block. Thus, 48 trials were included in total during the experiment. Examples of face stimuli are presented in Figure 1.

We used a Tobii TX300 eye tracker and the Tobii Studio software to present the stimuli, record eye movements, and analyze the gazing behavior of the participants. Figure 1 showed the stimuli used in the experiments. The fixation was defined as continuous gazing for more than 80 ms within a 1 degree of visual angle or 30 pixels. The experiment took place in a controlled environment (illumination, temperature, etc.) in the development laboratory of the Department of Maternal, Child & Adolescent Health at Tianjin Medical University.

Participants were instructed to look at the pictures on the monitor in a relaxed way. After completing the 9-point calibration, the test started with instruction text displayed on the screen explaining the procedure in detail. The pictures were presented in randomized order for duration of 5 s at a sampling rate of 120 Hz by using Tobii Studio 3.0 Eye Tracking Software. Each picture had a size of 720×480 pixels, subtending a visual angle of 13.78° in height by 7in width. When appearing together, the pictures were located approximately 5°visual angle away from each other. Between two trials, an image of a cartoon penguin over a white background was presented at the center of the screen for 1 s. While viewing the picture, the subject was not required to give a response. For uncooperative subjects, the caregiver was allowed to accompany the participant, but not to see the screen. A flow chart of the experiment is presented in Figure 1.

We used the entire area occupied by the emotional face as the Area of Interest (AOI). We analyzed the total fixation duration (TFD), i.e. the sum of the duration of the subject’s fixation in the AOI (24). Percent of total fixation duration to emotional vs. neutral faces were calculated, i.e. sum of fixation time per AOI was divided by the total sum of fixation time for both AOIs to derive the proportion of time spent on each AOI (i.e., “% emotional” and “%neutral”) and to correct for missing data (25).

A trained electrophysiologist delivered rTMS stimulation over the cortical area controlling the contralateral First Dorsal Interosseous (FDI) using a Magnetic Field Stimulator (CCY-1, YIRUIDE Medical Corporation, Wuhan, China) to detect resting motor threshold (MT). The MT was determined for each hemisphere in all individuals by gradually increasing the output of the machine by 5% until a 5 mV deflection or a visible twitch in the FDI muscle was identified in 2 out of 3 trials (26). Electromyographic (EMG) responses were monitored continuously from the hand contralateral to the stimulated hemisphere using the MEP module in Magnetic Stimulator (YIRUIDE Medical Corporation, Wuhan, China). Subjects were familiarized with the laboratory and procedure before the first TMS session. In this study, rTMS was selected to stimulate left DLPFC with high frequency (10 Hz) and right DLPFC with low frequency (1 Hz) based on the evidence-based basis proposed by the European Union of Neurological Societies (27), and the electrode positioning cap was used for accurate positioning. Specific parameters are as follows: stimulation frequency of right dorsolateral prefrontal lobe is 1Hz, stimulation time is 32s, stimulation number is 32, intermittent time is 1s, repetition number is 28, the stimulation intensity is 25% MT; stimulation frequency of left dorsolateral prefrontal lobe is 10Hz, stimulation time is 3.2s, stimulation number is 32. Intermittent time is 10s, repetition number is 45. Stimulation intensity is 25% MT.

LPA aims to identify types of people with different personal and/or environmental attributes profiles. In the domain of vocational behavior, frequently these personal and environmental attributes are psychological constructs (e.g. different types of commitment, different dimensions of career adaptability, and different types of perceived environmental support), so LPA can also be described as identifying construct-based profiles (28). We introduce a common LPA model equations as shown below:

where μ_ik and σ_ik represent profile-specific (k) means and variances for variable i, and π_k indicates profile density, or the proportion of N participants that belong to profile k. LPAs assume (a) samples drawn from a heterogeneous population produce data that are a mixture of K profile-specific distributions; (b) observed y indicator variables are distributed normally; and (c) the profile-specific mean vectors μ_k are the profile-specific (k) observed variable means.

LPA was conducted to investigate the optimal number of latent profiles that describe the patients’ fixation (mean TFD) on each of conditions of emotional faces. There were altogether 8 conditions, those were gender (male/female) × location(left/right visual field) ×valence (positive/negative). Models with two to seven profiles were tested. All analyses were conducted in Mplus 8.3 using robust maximum likelihood estimation. We used the following fit indicators to determine the optimal number of latent profiles (29, 30): Akaike information criterion (AIC), Bayesian information criterion (BIC), and adjusted Bayesian information criterion (aBIC) were employed to compare the model, with lower AIC, BIC, and aBIC values indicating a better model fit; the Lo-Mendell-Rubin (LMR) and bootstrap likelihood ratio test (BLRT) were examined to identify whether a k-class model fit better than a model with k-1 classes, and a significant p-value indicated that the k class was better; the entropy was assessed to identify each model’s classification precision with greater values implying more precise categorization (ideally above 0.80).

A regression mixture model (RMM) was used to analyze and compare the sex distribution, age, and group of individuals within the different profiles.

The dropout group consisted of fifty-six ASD participants, but five of them did not complete the eye-tracking experiments and three of them failed to finish the dropout questionnaire, and 48 remained at last, with a mean age of 7.995 (SD = 1.899; a range of 3.0–15.2) years. There were fifty-six ASD participants in the control group, with an average age of 8.802 (SD = 2.711; range of 3.3–17.8) years. There was no statistical difference (t = 1.080, P = 0.284) between the two groups. The sex distribution between the two groups was not statistically different either ( χ2=0.085, P = 0.771). Neither did the total score of CARS (38.96 ± 7.67 vs. 36.95 ± 6.82, t = 1.141, P = 0.258). see Table 1.

As shown in Table 2, the AIC, and aBIC generally decreased as the number of estimated profiles increased and reached the least values at Model 6, while the BIC stopped decreasing at Model 3. As to entropy, it stayed consistently above 0.80. According to LMR, the four-class solution did not enhance model fit considerably compared with the three-class solution (P = 0.4585). However, BLRT indicated that the six-class solution was no better than the five-class solution (P = 0.999). Based on model fit tests and the goal of parsimony, the three-class solution was identified as the best description of latent facial fixation profiles. The latent profile memberships showed significant differences in the means of the eight indicator variables (Table 3), and their characteristics are summarized in Figure 2. Class 1 (n = 84, 80.76%), the moderate undeferential fixation group, was characterized by the average fixations on each condition without difference. Class 2 (n = 10, 9.62%), the less but diverse fixation group, was distinguished by a relatively less fixation but with intensive variance. Class 3 (n = 10, 9.62%), the more fixation with mild change group, was characterized by the highest fixation on emotional faces in most conditions.

In addition, there was no statistical difference among the three latent profiles in terms of age, but not for sex distribution. However, the total score of CARS was statistically significant, but only between the “moderate undeferential fixation group” and “more fixation with mild change group”, the total CARS scores were significantly higher in the “moderate undeferential fixation group”, see Table 4. LPA as a predictor of dropout

The RMM revealed significant differences between the three profiles regarding the group (χ² = 18.458, P <0.001), but only between the “moderate undeferential fixation group” and “more with mild change group”. The intolerance rate in the three profiles was 38.1%, 60.0%, and 100%.And the latent profile was the influential factor of dropout, with β = -1.559,S.E = 0.497, Wald = 9.830, P = 0.002, OR = 0.210 (95%CI: 0.079, 0.557), which indicated that the participants in “less but diverse fixation group” or “more fixation with mild change group” are more easily to drop out than those in “moderate undeferential fixation group”.