MindTS-Net: A multimodal mindfulness-based intervention framework for children with Tourette syndrome

Summary

Tourette syndrome affects children’s daily functioning and creates a need for supervised support that responds to tic behaviors, emotional states, and interaction contexts. We developed MindTS-MMD, a Chinese multimodal dataset capturing emotion, tic-related behaviors, interaction context, and Valence-Arousal ratings, and MindTS-Net, a supervised digital adjunct linking multimodal emotion perception, mindfulness strategy planning, child-appropriate guidance generation, and dialogue-state tracking. Evaluation included dataset benchmarking, a 14-day primary clinical study with 60 children, and an independent external validation cohort of 57 children from a second clinical center. MindTS-Net improved emotion perception, strategy selection, and guidance quality compared with baseline approaches. In the primary study, children receiving MindTS-Net showed favorable Valence-Arousal changes and reduced tic burden. External validation showed consistent directions across perception, strategy, guidance, and intervention outcomes. These findings suggest that supervised multimodal digital support can complement pediatric tic management and provide a foundation for future multi-center research.

Graphical abstract

Highlights

• •MindTS-Net integrates multimodal emotion perception for supervised TS support
• •Cross-modal emotion modeling supports adaptive selection of mindfulness strategies
• •Dialogue-state tracking enables child-appropriate mindfulness guidance
• •Clinical studies supported preliminary cross-center consistency

Health sciences

Introduction

Tourette syndrome is a neuropsychiatric condition that typically emerges in childhood and is characterized by involuntary and repetitive motor or vocal tics, frequently accompanied by comorbidities such as attention-deficit/hyperactivity disorder (ADHD) and anxiety disorders.¹,² Heightened emotional arousal and affective instability have been reported as proximal triggers of tic exacerbation, although long-term outcomes are shaped by multiple factors including the waxing-and-waning course of tics, baseline severity, comorbidities, and access to evidence-based care.³ Consequently, developing child-centered interventions that support emotion regulation is of substantial clinical and societal significance. Mindfulness-based interventions (MBIs) cultivate present-moment awareness and non-judgmental acceptance and have been used as adjunctive therapies for childhood anxiety, impulse-control difficulties, and attentional deficits.⁴ Recent advances in large language models (LLMs) have further opened opportunities for intelligent, interactive, and dynamically adaptive psychological support.⁵ However, existing affective dialogue systems mainly provide empathy and comfort while lacking goal-oriented intervention planning, semantic controllability, and adaptation to children’s developmental and cognitive profiles.⁶

Children with TS often present impairments in emotion recognition, expression, and regulation, frequently accompanied by impulsivity.⁷ Anxiety is more prevalent in this population and correlates positively with tic frequency,⁸ while comorbid ADHD can further exacerbate deficits in attentional and emotional control.⁹ Together, these factors may create a maladaptive cycle in which dysregulation both intensifies and is aggravated by tic severity.¹⁰ In current clinical practice, medication can alleviate tics but often provides limited emotional benefits and may be associated with side effects.^11,12 Clinical guidelines therefore recommend cognitive behavioral therapy (CBT)¹³ and habit reversal training (HRT)¹⁴ as first-line interventions. Comprehensive behavioral intervention for tics (CBITs),¹⁵ which integrates core HRT elements with psychoeducation, functional intervention, and relapse prevention, has demonstrated efficacy in randomized trials,¹⁶ and developmentally adapted formats such as CBIT-JR support the feasibility of behavioral treatment in younger children.¹⁷ However, these approaches remain therapist-intensive, usually require repeated in-person sessions, and are difficult to sustain in daily life. In this context, mindfulness-based strategies are better viewed as complementary to evidence-based behavioral treatments such as CBIT/HRT, because they primarily target emotional arousal and attentional control—proximal factors that often precede tic exacerbation in daily life.¹⁸ Behavioral studies further suggest that structured interventions can reduce tic frequency and improve emotional regulation.^19,20,21 Neurobiological evidence has also implicated prefrontal-striatal abnormalities and Treg dysregulation in impaired inhibition, stress reactivity, and symptom fluctuation.^22,23,24 The present study does not target immune mechanisms directly but instead focuses on modulating emotional arousal and stress reactivity as modifiable proximal triggers of tic exacerbation.²⁵ Family and school environments also matter: positive parenting and emotional support promote resilience, whereas adversity, family conflict, and social exclusion may worsen symptoms and adjustment.^{26,27,28,29,30}

Mindfulness-based methods have been explored to address these challenges. Evidence suggests that MBCT, CBT, and ACT can enhance emotion regulation and reduce symptom severity in related conditions.³¹ Digital MBIs such as headspace and calm,³² mindfulness-based music interventions,³³ and related studies^34,35 have broadened access, but real-time adaptability and deep therapeutic engagement remain limited. Researchers have therefore begun integrating multimodal perception and intelligent dialogue into intervention systems. Adaptive graph attention improves EEG-based emotion recognition,³⁶ multi-source adaptive fusion enhances cross-domain robustness,³⁷ interactive attention strengthens multimodal speech-text recognition,³⁸ and multimodal physiological signals further improve emotional state identification.³⁹ Related advances in multimodal datasets, fusion architectures, affective sensing, and dialogue modeling^{40,41,42,43,44} further support the feasibility of such systems. Despite these advances, most existing systems target adults or general contexts rather than children,⁴⁵ multimodal corpora specifically designed for pediatric interventions remain scarce,⁴⁶ and perception-strategy coupling, continuous state tracking, and mindfulness-oriented controllable generation are still insufficient.^47,48,49 Emerging reviews further suggest that mindfulness-based approaches may offer potential benefits for Tourette syndrome and tic disorders, particularly in symptom management, emotional regulation, and treatment accessibility, although dedicated pediatric evidence remains limited.⁵⁰

Taken together, these gaps point to the need for a supervised, child-centered digital support system that can recognize emotional states, provide developmentally appropriate regulation guidance, and assist daily monitoring beyond clinic-based sessions.⁵¹ The aim of this work is not to replace first-line behavioral treatment, provide autonomous diagnosis, recommend medication, or digitize CBIT as a fully automated intervention. Rather, we focus on a complementary and scalable between-visit support layer that regulates emotional arousal and supports context-sensitive monitoring under caregiver and researcher supervision. To address this need, we propose MindTS-Net, a supervised multimodal digital adjunct supported by MindTS-MMD, a multimodal dataset specifically designed for pediatric TS intervention research. MindTS-Net links multimodal emotion perception, mindfulness strategy planning, age-appropriate guidance generation, and dialogue-state tracking within a supervised intervention workflow. Detailed model architecture, training objectives, hyperparameters, and implementation procedures are provided in STAR Methods. The contributions of this work are summarized as follows.

• •We construct MindTS-MMD, a clinically annotated multimodal dataset for pediatric TS intervention research, which synchronously captures emotional states, tic-related behaviors, interaction context, and Valence-Arousal ratings, with high inter-rater reliability (Cohen’s κ = 0.83).
• •We introduce MindTS-Net as a supervised multimodal digital adjunct for pediatric TS, integrating multimodal emotion perception, mindfulness strategy planning, child-appropriate guidance generation, and dialogue-state tracking to support personalized emotion regulation.
• •We implement a supervised perception-strategy-intervention workflow that converts multimodal emotional and contextual information into developmentally appropriate mindfulness guidance under predefined safety constraints and caregiver/researcher supervision.
• •We conduct dataset-based benchmarking, a 14-day clinically supervised study, and an independent external validation study. Benchmark experiments provide preliminary statistical evidence of improved emotion recognition and expert-rated guidance appropriateness, while participant-level clinical analyses and external validation results support preliminary cross-center consistency of the proposed intervention pipeline.

Results

The following analyses are based on the MindTS-MMD dataset and the MindTS-Net framework. MindTS-MMD provided synchronized multimodal interaction data with emotion, Valence-Arousal, and tic-related annotations, while MindTS-Net connected emotion perception, strategy planning, guidance generation, and dialogue-state tracking for supervised pediatric TS intervention. Details on dataset construction, cohort recruitment, model architecture, training, and evaluation are provided in STAR Methods.

Multimodal emotion perception

We first evaluated MindTS-Net on MindTS-MMD to examine whether multimodal signals improved the recognition of clinically relevant emotional states in children with TS. The full text-audio-visual configuration (T + A + V) was compared with the text-audio baseline (T + A) for both discrete emotion classification and continuous Valence-Arousal modeling.

As shown in Table 1, the integration of visual input improved performance across all seven discrete emotion categories. The largest gains were observed in categories characterized by sparse or ambiguous verbal expression, particularly Tense and Shy. For the Tense category, F1-score increased from 53.1% to 62.7%, while Shy improved from 48.4% to 60.0%. These findings suggest that visual cues provide clinically useful complementary information when children’s verbal or acoustic emotional cues are limited or ambiguous.

Table 1.

Performance comparison of emotion classification by category (%)

Emotion category	Precision (T + A)	Precision (T + A + V)	Recall (T + A)	Recall (T + A + V)	F1-score (T + A)	F1-score (T + A + V)
Calm	79.3	84.1	82.5	86.2	80.9	85.1
Anxious	70.4	73.2	69.3	70.1	69.8	71.6
Irritable	66.1	70.5	64	67.8	65.0	69.0
Tense	60.3	66.9	47.5	59.1	53.1	62.7
Relaxed	75.4	79.8	72.2	76.5	73.8	78.1
Shy	58.6	66.3	41.2	54.8	48.4	60.0
Focused	69.1	74.2	66.8	71.5	67.9	72.8

The confusion matrices in Figure 1 further show that visual information reduced misclassifications among low-salience emotional states. In particular, the T + A + V configuration showed clearer separation for Shy/Calm and Tense/Anxious than the T + A baseline. These results indicate that facial expression, gaze, and posture may help capture subtle emotional states that are difficult to infer from text and audio alone.

Figure 1.

Confusion matrices comparing T + A and T + A + V

(A) Confusion matrix for T+A.

(B) Confusion matrix for T+A+V.

For continuous emotion modeling, the T + A + V configuration aligned more closely with ground-truth Valence-Arousal labels than the T + A baseline. As shown in Figure 2, predicted Valence and Arousal values were closer to the reference diagonal under the multimodal configuration.

Figure 2.

Predicted vs. ground-truth values for Valence and Arousal

(A) Valence prediction.

(B) Arousal prediction.

Table 2 further shows that the T + A + V model achieved higher PCC values for both Valence and Arousal and lower MSE values compared with the T + A baseline. These results indicate that multimodal perception improved both categorical emotion recognition and dimensional emotional-state estimation.

Table 2.

Regression metrics for Valence and Arousal

Modality (metric)	Mean ± SD	SE	95% CI	t	p value
T + A (Valence-PCC)	0.741 ± 0.024	0.014	[0.714, 0.768]	6.014	0.003
T + A + V (Valence-PCC)	0.809 ± 0.008	0.005	[0.799, 0.819]	–	–
T + A (Arousal-PCC)	0.690 ± 0.019	0.011	[0.668, 0.712]	5.732	0.004
T + A + V (Arousal-PCC)	0.780 ± 0.013	0.008	[0.764, 0.796]	–	–
T + A (Valence-MSE)	0.146 ± 0.016	0.009	[0.128, 0.164]	4.877	0.008
T + A + V (Valence-MSE)	0.128 ± 0.006	0.003	[0.122, 0.134]	–	–
T + A (Arousal-MSE)	0.161 ± 0.022	0.013	[0.136, 0.186]	5.143	0.006
T + A + V (Arousal-MSE)	0.143 ± 0.011	0.006	[0.131, 0.155]	–	–

We further examined whether tic manifestations affected emotion perception. Samples were divided into low- and high-tic subsets using a median split of annotated tic frequency within each observation window. As shown in Table 3, emotion perception performance decreased under high tic frequency, with reductions in both overall accuracy and Macro-F1. The performance drop was mainly associated with low-salience emotional states, particularly Shy/Calm and Tense/Anxious. These findings suggest that tic-related motor and vocal bursts may overlap with subtle emotional cues, increasing the difficulty of emotion recognition in children with TS.

Table 3.

Tic-frequency-stratified evaluation of emotion perception performance

Assessment dimension	Low-tic subset	High-tic subset	Interpretation
Overall accuracy	0.783	0.741	recognition accuracy decreased under high tic frequency
Macro-level F1	0.724	0.673	balanced emotion perception was reduced under high tic frequency
Main confusion pattern	lower confusion among low-salience states	increased confusion among Shy/Calm and Tense/Anxious	tic-related cues may interfere with subtle emotional signals
Evaluation protocol	full T + A + V setting	full T + A + V setting	stratified evaluation based on median-split tic frequency

MindTS-Net generates clinically acceptable mindfulness strategies and guidance

We next evaluated whether MindTS-Net could translate recognized emotional states into appropriate mindfulness strategies and child-centered guidance. In closed-set strategy classification, MindTS-Net achieved higher strategy-selection accuracy than rule-based and text-only baselines. In open-set semantic matching, generated strategy representations showed stronger alignment with expert-authored references. These findings suggest that multimodal emotional information supported the selection of clinically meaningful mindfulness strategies.

MindTS-Net also produced guidance responses with higher semantic consistency, mindfulness-oriented style, and linguistic diversity than comparison systems. Automated metrics, including BLEU, ROUGE-L, BERT Score, cosine similarity, and Distinct scores, favored MindTS-Net. Subjective ratings by five professional counselors further supported the naturalness, mindfulness style, contextual consistency, and child-oriented appropriateness of the generated responses. Detailed benchmark comparisons for strategy classification, semantic matching, and language-generation quality are provided in Figure S2 and Tables S4–S6.

Representative examples further illustrate the clinical interpretability of the generated guidance. Compared with generic mindfulness-based instructions, MindTS-Net produced responses conditioned on the recognized emotional state and adapted to children’s cognitive and linguistic characteristics, as shown in Table 4.

Table 4.

Examples of traditional mindfulness guidance and MindTS-Net responses

Emotion state	Traditional MBI guidance	MindTS-Net generated guidance
Calm	focus on your breathing and stay present in the moment.	you seem quite calm right now. Let’s take a gentle breath and simply notice how your body feels while everything is quiet and steady.
Relaxed	continue breathing slowly and maintain a relaxed state.	your body looks relaxed. Let’s enjoy this calm moment together—take a slow breath and notice how comfortable your shoulders and arms feel.
Focused	keep your attention on the present moment and your breathing.	you seem really focused. Let’s keep that attention on your breathing for a few seconds and notice each slow inhale and exhale.
Shy	try to stay calm and focus on the present moment.	if you feel a little shy right now, that’s okay. Let’s take a quiet breath and notice something calm around you, such as the feeling of your feet on the ground.
Anxious	take a few deep breaths and try to relax.	let’s take a slow breath together. As you breathe out, imagine your shoulders getting lighter and your body becoming calmer.
Tense	try to calm down and focus on your breathing.	i can see your body might feel a little tight right now. Let’s gently breathe in and slowly relax your shoulders and hands.
Irritable	pause and breathe slowly until you feel calmer.	it’s okay to feel a little upset. Let’s pause for a moment and take two slow breaths together, letting the feeling pass such as a small cloud.

Notes: Original responses were generated in Mandarin Chinese and translated into English for presentation.

Primary clinical evaluation

The primary clinical evaluation examined whether MindTS-Net use was associated with changes in emotion regulation and observed tic frequency during the 14-day supervised protocol. Given the pilot design, clinical outcome analyses were interpreted as preliminary effect signals rather than definitive estimates of efficacy. Repeated daily observations were aggregated at the participant level before inferential testing.

During the intervention phase, the MindTS-Net group showed a clearer shift toward more positive and stable emotional states than the monitoring-only control group. As shown in Figure 3, the MindTS-Net group displayed a steady increase in Valence and a synchronous decrease in Arousal starting from approximately Day 5, whereas the control group showed only minor fluctuations without a clear directional trend.

Figure 3.

Emotion score trends across groups

Tic analysis based on standardized video observation windows showed a reduction in observed daily tic counts in both groups. However, the reduction was numerically larger in the MindTS-Net group, with mean observed daily tic counts decreasing from 18.500 to 4.012, whereas the control group decreased from 18.400 to 6.874. As shown in Figure 4, the MindTS-Net group exhibited a rapid initial decline followed by relatively stable maintenance, whereas the control group showed a more gradual and fluctuating trajectory. Participant-level change-score analysis further showed a between-group difference in mean change of 3.024 events per day in favor of MindTS-Net.

Figure 4.

Tic frequency trends over study phases

To provide an exploratory view of how system-assigned strategies related to short-term emotional changes, outcomes were summarized by dominant strategy category. As shown in Table 5, breathing regulation (S1) and awareness guidance (S6) were associated with the largest average increases in ΔValence (+2.124 and +2.034, respectively) and the largest decreases in ΔArousal (−1.542 and −1.482, respectively). In contrast, emotion labeling (S3) showed a comparatively smaller mean increase in ΔValence (+1.473) and decrease in ΔArousal (−1.026). Because strategies were adaptively assigned and individual participants could contribute to multiple strategy categories, these findings should be interpreted as exploratory and non-causal rather than as evidence of differential strategy efficacy.

Table 5.

Exploratory descriptive analysis of emotional changes across strategy categories

Strategy	N (Obs.)	ΔValence (mean ± SD)	95% CI [ΔV]	ΔArousal (mean ± SD)
S1	412	+2.124 ± 0.712	[1.861, 2.387]	−1.542 ± 0.521
S2	128	+1.892 ± 0.824	[1.587, 2.197]	−1.315 ± 0.614
S3	364	+1.473 ± 0.731	[1.203, 1.743]	−1.026 ± 0.582
S4	162	+1.621 ± 0.814	[1.321, 1.921]	−1.144 ± 0.635
S5	148	+1.565 ± 0.882	[1.239, 1.891]	−1.218 ± 0.672
S6	109	+2.034 ± 0.763	[1.752, 2.316]	−1.482 ± 0.546

Note. Values are presented as descriptive participant-level summaries by strategy category. Because strategies were adaptively assigned during the intervention and individual participants could contribute to multiple strategy categories, these results should be interpreted as exploratory and non-causal.

We next examined whether improvements in emotional regulation were associated with reductions in tic frequency. Participant-level summaries from the post-intervention phase were used to assess the correlation between emotional alignment improvement, defined as a shift toward higher Valence and lower Arousal, and tic frequency reduction rate (TFRR). Pearson’s correlation analysis showed a significant positive association in the MindTS-Net group (r = 0.595, p < 0.001; R² = 0.353), indicating that participants with greater emotional alignment improvement tended to show larger reductions in tic frequency. This relationship is shown in Figure 5. Because this analysis was correlational, it should not be interpreted as evidence of a direct causal effect.

Figure 5.

Correlation between emotional alignment improvement and tic frequency reduction

To evaluate whether MindTS-Net provided flexible and personalized intervention support, we analyzed strategy allocation across tic-frequency states and individual users. Participants were divided into high- and low-frequency tic groups using a median split. As shown in Table 6, strategy allocation differed between the two groups. The high-frequency tic group received higher proportions of breathing regulation (S1: 31.242%) and emotion labeling (S3: 27.186%) than the low-frequency group, with mean differences of 15.558 and 10.759 percentage points, respectively. This pattern suggests that MindTS-Net tended to prioritize arousal reduction and emotional grounding when tic burden was higher. In contrast, the low-frequency group received more relaxation training (S5: 23.864%) and awareness guidance (S6: 22.131%), indicating a shift toward maintenance-oriented and awareness-based strategies under lower symptom frequency. Body scan (S2) and mindfulness attention (S4) showed no significant between-group differences, suggesting that these strategies functioned as stable supportive components across symptom states.

Table 6.

Mean participant-level strategy usage rates in the high- and low-frequency tic groups

Strategy	High-freq (%) ± SD	Low-Freq (%) ± SD	Mean Diff	95% CI	p value
S1	31.242 ± 5.381	15.684 ± 4.572	15.558	[12.412, 18.704]	<0.001
S2	09.815 ± 3.124	08.632 ± 2.756	1.183	[−1.224, 3.590]	0.324
S3	27.186 ± 4.762	16.427 ± 3.915	10.759	[8.115, 13.403]	<0.001
S4	12.351 ± 3.648	13.264 ± 4.213	−0.913	[−3.150, 1.324]	0.412
S5	11.084 ± 2.937	23.864 ± 5.068	−12.78	[−15.221, −10.339]	<0.001
S6	08.322 ± 2.415	22.131 ± 4.832	−13.809	[−16.042, −11.576]	<0.001

At the individual level, strategy allocation showed substantial heterogeneity across participants. The mean entropy of the strategy distribution was 1.914, approaching the theoretical maximum of 2.585 for six strategies. These results suggest that MindTS-Net did not rely on a fixed template but adjusted intervention strategies according to both tic-frequency state and individual response patterns. The aggregated strategy allocation heatmap is provided in Figure S4.

External validation supports cross-center consistency

To evaluate cross-center consistency, we compared the primary MindTS-Net group with an independent external validation cohort from the Second Affiliated Hospital of Xinxiang Medical University. Because the external cohort received MindTS-Net-guided intervention in addition to conventional medication and did not include a separate control group, this analysis was interpreted as cross-center consistency rather than controlled efficacy.

Emotion perception showed only minor fluctuations in the external cohort compared with the primary MindTS-Net group. Accuracy decreased slightly, whereas Macro-F1 showed a small numerical increase. Continuous Valence-Arousal estimation also remained stable, with small bidirectional changes in PCC and MSE values. None of these differences reached statistical significance, and effect sizes remained small. These findings suggest that the emotion perception component retained stable performance under a second clinical setting.

Strategy selection and guidance generation also showed no statistically significant cross-center degradation. Strategy accuracy, strategy consistency, semantic matching, language-generation metrics, and expert ratings were broadly comparable between cohorts. These results indicate that the perception-strategy-guidance pathway retained functional stability and clinically acceptable output quality in the external cohort. Detailed cross-center comparisons of emotion perception, strategy selection, and guidance generation are provided in Figure S5, Tables S9 and S10.

The most clinically relevant comparison concerned intervention-related outcomes. As shown in Table 7, final Valence and Arousal scores were similar between the primary MindTS-Net group and the external validation cohort. Daily tic count was slightly higher in the external cohort, but the difference was not statistically significant. TFRR remained high in both cohorts, reaching 78.412% internally and 75.144% externally. Completion rates were also high and comparable across cohorts, supporting the feasibility and adherence of the supervised intervention workflow across clinical settings.

Table 7.

Comparison of intervention effects in internal cohort and external validation cohort

Metric	Internal (n = 30)	External (n = 57)	Mean Diff	95% CI	Cohen’s d	p value
Valence score	2.364 ± 0.942	2.291 ± 1.046	0.073	[−0.354, 0.500]	0.073	0.736
Arousal score	1.342 ± 0.815	1.417 ± 0.948	−0.075	[−0.451, 0.301]	0.085	0.694
Daily Tic count	4.012 ± 3.184	4.658 ± 3.762	−0.646	[−2.152, 0.860]	0.185	0.398
TFRR (%)	78.412 ± 18.423	75.144 ± 21.867	3.268	[−5.402, 11.938]	0.162	0.457
Completion %	94.152 ± 8.412	91.537 ± 9.746	2.615	[−1.342, 6.572]	0.287	0.192

Taken together, the external validation results showed consistent outcome directions across emotion perception, strategy selection, guidance generation, and intervention-related indicators. These findings support preliminary cross-center consistency of the supervised MindTS-Net pipeline, while larger multi-center studies with active control conditions and longer follow-up are still required to confirm efficacy and long-term generalizability.

Discussion

MindTS-Net provides a supervised multimodal digital adjunct for children with Tourette syndrome by linking emotion perception, mindfulness strategy planning, child-appropriate guidance generation, and longitudinal state tracking within one intervention workflow. The central finding of this study is clinical rather than purely technical: during the 14-day supervised protocol, children in the MindTS-Net group showed a clearer shift toward higher Valence and lower Arousal than the monitoring-only control group, together with a larger numerical reduction in observed daily tic counts. These findings suggest that a child-centered digital support system targeting emotional arousal and daily regulation may provide useful between-visit support for children with TS. Given the pilot design, these results should be interpreted as preliminary clinical signals rather than definitive evidence of efficacy.

The emotional changes observed in this study are clinically meaningful because tic symptoms are often sensitive to affective arousal, stress, and contextual triggers. MindTS-Net does not directly retrain tic behaviors in the same manner as CBIT or HRT; instead, it focuses on emotion regulation, arousal stabilization, and context-sensitive mindfulness guidance. The positive association between emotional alignment improvement and tic-frequency reduction further supports this interpretation. Children who showed greater movement toward higher Valence and lower Arousal tended to show larger reductions in tic frequency. However, this relationship was correlational and should not be interpreted as evidence that emotional improvement directly caused tic reduction. Rather, it suggests that emotional regulation and tic burden may change together during supervised digital intervention.

The strategy-level findings provide additional insight into how MindTS-Net may support short-term regulation. Breathing regulation and awareness guidance were associated with larger average improvements in Valence and Arousal than some other strategies, while strategy allocation differed between high- and low-frequency tic states. In higher tic-frequency states, the system tended to assign more breathing regulation and emotion labeling, whereas lower tic-frequency states were more often associated with relaxation training and awareness guidance. This pattern suggests that MindTS-Net did not rely on a single fixed template but adjusted support according to symptom state and individual response patterns. Because strategies were adaptively assigned rather than randomized, these results should be viewed as descriptive and hypothesis-generating rather than as evidence of differential strategy efficacy.

Multimodal emotion perception contributed to the clinical usability of the system. Children with TS may not always express emotional distress through clear verbal reports, and subtle facial, vocal, or behavioral cues may provide important complementary information. In this study, adding visual input improved recognition of low-salience emotional states such as shyness and tension, which are clinically relevant but often difficult to infer from text and audio alone. This suggests that multimodal sensing can help detect affective states that children may not clearly verbalize, thereby supporting more context-sensitive intervention planning. In this respect, the value of multimodal perception lies not only in benchmark performance but also in its potential to strengthen the system’s responsiveness to clinically relevant emotional states during supervised interactions.

The external validation cohort further supports the stability of the proposed intervention pipeline across clinical settings. Compared with the primary MindTS-Net group, the external cohort showed broadly comparable patterns in emotion perception, strategy selection, guidance generation, and intervention-related outcomes. Final Valence, Arousal, daily tic count, TFRR, and completion rate showed no statistically significant cross-center differences, and the overall direction of change was consistent. Because the external cohort did not include a separate control group, these findings should be interpreted as preliminary cross-center consistency rather than controlled external efficacy. Nevertheless, the results suggest that the supervised perception-strategy-intervention workflow was not limited to the original clinical setting.

From a clinical perspective, MindTS-Net should be understood as a complementary digital support tool rather than a replacement for first-line behavioral treatments such as CBIT or HRT. CBIT and HRT directly target tic behaviors through structured competing response training, functional analysis, and therapist-guided behavioral practice. By contrast, MindTS-Net focuses on emotion regulation, arousal reduction, and context-sensitive support between clinic visits. This distinction is important because the system is not intended to provide autonomous diagnosis, medication advice, or fully automated behavioral therapy. Its potential role is to extend supervised support into daily contexts where therapist-intensive intervention is not continuously available.

Together, these findings support MindTS-Net as a scalable, supervised, and clinically bounded digital adjunct for pediatric tic management. By combining child-centered multimodal emotion perception with mindfulness strategy planning and adaptive guidance, the framework provides preliminary evidence for supporting short-term emotional regulation and tic-related monitoring in daily contexts. The results also establish MindTS-MMD as a useful multimodal resource for pediatric TS intervention research. Further validation will be required before broad clinical deployment, but the present findings provide a foundation for larger controlled studies of AI-assisted, emotion-focused support for children with TS.

Limitations of the study

Several limitations should be considered. First, although this study included an independent external validation cohort, the validation was conducted in only one additional hospital, and the external cohort did not include a separate control group. Therefore, the external results should be interpreted as evidence of cross-center consistency rather than controlled external efficacy. Second, the intervention period was limited to 14 days, which is short relative to the waxing-and-waning course of tic symptoms; longer follow-up is needed to determine whether the observed improvements in Valence, Arousal, and daily tic counts can be sustained over time. Third, the primary control group received monitoring only rather than an engagement-matched active digital control, so expectancy effects, caregiver prompting, and adherence-related influences cannot be fully excluded. Finally, strategy-level analyses were exploratory because strategies were adaptively assigned rather than randomized, and MindTS-Net is currently optimized for Mandarin Chinese and the cultural-pragmatic context represented in MindTS-MMD. Future studies should include larger multi-center cohorts, longer follow-up, active control conditions, and culturally adapted validation to further establish the long-term clinical utility and real-world deploy ability of this supervised digital adjunct.

By demonstrating preliminary improvements in emotion regulation and tic-related outcomes under supervised use, this study provides a foundation for future controlled trials of multimodal, emotion-focused digital support for children with Tourette syndrome.

Resource availability

Lead contact

Further information and requests for resources should be directed to Fang Xiaoxia (xxchfangxiaoxia@163.com).

Materials availability

This study did not generate new physical materials.

Data and code availability

• •Dataset: Only de-identified derived representations, annotations, and selected non-identifiable features from the MindTS-MMD multimodal pediatric Tourette syndrome dataset, together with aggregated external validation descriptors, are publicly available through Zenodo. The accession no. (https://doi.org/10.5281/zenodo.18281896) is listed in the key resources table. Raw audio and video recordings from the primary clinical evaluation cohort and the external validation cohort are not publicly available because of participant privacy and ethical restrictions. Any sharing of raw recordings, if permitted, would require controlled access, approval from the relevant ethics committees, and compliance with applicable data-use agreements.
• •Code: The MindTS-Net training and inference code (version 1.0), including training configurations and evaluation scripts used in this study, is publicly available at https://github.com/wangjianpinghist-ctrl/MindTS-Net, as listed in the key resources table. Configuration files required to reproduce the reported analyses are included in the repository.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon reasonable request and, where applicable, subject to ethics approval and institutional data-sharing regulations.

Acknowledgments

This work was supported in part by the Henan Provincial Science and Technology Research Project under grant no. 252102211022. We sincerely thank all children and families who participated in the study, as well as the clinicians and research staff at Xinxiang Central Hospital and the Second Affiliated Hospital of Xinxiang Medical University for their assistance with participant recruitment, data collection, clinical coordination, and external validation.

Author contributions

Conceptualization and design: L.L. and X.F.; methodology: L.L., J.W., and X.F.; data collection and preprocessing: H.Z. and H.S.; model implementation, validation, and technical optimization: J.W., H.W., and Q.L.; external validation analysis: J.W., H.W., Q.L., and X.F.; statistical analysis: L.L.; clinical interpretation and discussion: L.L., H.S., and X.F.; manuscript drafting: L.L.; manuscript review and editing: all authors; project administration: L.L.; supervision: X.F.; funding acquisition: L.L.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Biological samples
Human participants: Primary clinical evaluation cohort	This paper	Xinxiang Central Hospital; n = 60
Human participants: External validation cohort	This paper	Second Affiliated Hospital of Xinxiang Medical University; n = 57
Deposited data
MindTS-MMD (multimodal pediatric TS dataset)	This paper; Zenodo	https://doi.org/10.5281/zenodo.18281896
Software and algorithms
MindTS-Net (training and inference code, v1.0)	This paper	https://github.com/wangjianpinghist-ctrl/MindTS-Net
Chinese-BERT-WWM	HuggingFace	https://huggingface.co/hfl/chinese-bert-wwm
Whisper (automatic speech recognition)	OpenAI	https://github.com/openai/whisper
COVAREP v1.5	Degottex et al.	https://github.com/covarep/covarep
OpenFace v2.2	Baltrušaitis et al.	https://github.com/TadasBaltrusaitis/OpenFace
PyTorch v2.0.1	PyTorch Foundation	https://pytorch.org
Other
Ubuntu 20.04 LTS	Canonical	https://ubuntu.com
NVIDIA A100 GPUs	NVIDIA	40–80 GB memory

Experimental model and study participant details

This prospective supervised-intervention study enrolled Chinese children aged 6–12 years with clinician-confirmed Tourette syndrome, recruited from Xinxiang Central Hospital and the Second Affiliated Hospital of Xinxiang Medical University. Diagnosis was established according to DSM-5 criteria through structured clinical interviews and medical record review. No restrictions were applied based on biological sex, social gender, ancestry, race, or ethnicity.

Primary clinical evaluation cohort

At Xinxiang Central Hospital, 74 children were assessed for eligibility. After screening, 65 participants were randomized using a computer-generated allocation sequence in a 1:1 ratio. Sixty participants completed the study and were included in the final participant-level analysis, with 30 assigned to the MindTS-Net intervention group and 30 assigned to the monitoring-only control group. The study procedure consisted of a 14-day protocol including a 4-day baseline observation phase followed by a 10-day intervention phase.

External validation cohort

To evaluate cross-center consistency, an independent cohort was established at the Second Affiliated Hospital of Xinxiang Medical University. A total of 63 children were recruited and followed the same 14-day protocol as the primary cohort. All participants in this cohort received MindTS-Net-guided intervention in addition to conventional medication. Six participants were excluded prior to statistical analysis due to predefined data-quality criteria, leaving 57 participants for the final analysis. As in the primary cohort, repeated daily observations were aggregated at the participant level before outcome comparisons. Data from the external cohort were used exclusively for independent validation and were not used for model training, hyperparameter tuning, or selection.

Ethical considerations

The study protocol was approved by the Medical Ethics Committee of Xinxiang Central Hospital (approval no. 2022-114-01(K)) and the Ethics Committee of the Second Affiliated Hospital of Xinxiang Medical University (approval no. URG-2026023). Written informed consent was obtained from the legal guardians of all participants prior to enrollment, and age-appropriate assent was obtained from children whenever applicable. The study was registered with the Chinese Clinical Trial Registry (ChiCTR2500102153) and conducted in accordance with institutional ethical requirements, national regulations, and the Declaration of Helsinki. All multimodal recordings were de-identified before annotation and analysis, and personally identifiable information was removed from the research dataset.

Sex and gender considerations

Both male and female participants were included. Biological sex was recorded during enrollment, but the study was not specifically powered to detect sex-based differences in outcomes, and no sex-stratified analyses were performed. Social gender-related factors were not systematically collected. Therefore, the potential influence of sex and gender on observed findings cannot be fully assessed and represents a limitation of the study.

Ancestry, race, and ethnicity

No ancestry- or race-based eligibility criteria were applied, and ancestry, race, and ethnicity data were not systematically collected.

Method details

MindTS-MMD dataset construction

MindTS-MMD was constructed as a Chinese, child-centered multimodal dataset for pediatric Tourette syndrome intervention research. The dataset was designed to support supervised modeling of multimodal emotion perception, tic-related behavioral annotation, mindfulness strategy planning, and child-appropriate guidance generation. Existing multimodal corpora and facial-expression datasets do not fully meet the requirements of synchronized text–audio–video interaction data, tic-related annotations, and mindfulness-oriented pediatric intervention modeling; therefore, MindTS-MMD was developed specifically for this study.

Data were collected from children aged 6–12 years using a semi-structured emotion induction paradigm, including picture stimuli, story retelling, situational narration, and role-playing tasks. Each session contained 6–10 interaction rounds organized around six predefined mindfulness-based intervention strategies, while emotion labels were annotated independently from strategy labels. Task order was pseudo-randomized, and neutral or low-arousal segments were inserted between high-arousal conditions to reduce emotional carryover.

Each interaction round was designed to capture synchronized text, audio, and video data. Text data were obtained from Mandarin ASR transcripts, audio was recorded as 16 kHz WAV files, and video was recorded in H.264 MP4 format at 25–30 frames per second. Whisper was used for Mandarin ASR and timestamp alignment. COVAREP and OpenFace were used for acoustic and visual preprocessing, auxiliary feature extraction, and quality control. Word-level timestamps were used to align audio and video streams within a tolerance of ±120 ms, and segments exceeding this tolerance were excluded from synchronized multimodal analysis.

After preprocessing and quality control, synchronized multimodal records were annotated for emotion category, Valence–Arousal scores, tic frequency, and tic location. Seven elicited emotion categories were used: calm, relaxed, focused, shy, anxious, tense, and irritable. Valence–Arousal scores were obtained using a dual-source rating scheme combining child self-reports and experimenter observations, with both dimensions rated on a 9-point scale from −4 to +4. Inter-rater reliability for categorical emotion annotation reached Cohen’s κ = 0.83. Tic-related annotations were obtained through synchronized audio–video review by trained raters and were linked to the corresponding interaction windows for downstream modeling and clinical analysis.

The final fully labeled multimodal analysis set contained 5,498 rounds with valid video recordings and synchronized annotations for emotion category, Valence–Arousal scores, tic frequency, and tic location. Additional audio-only or text-available rounds retained during preprocessing were not included in the fully supervised multimodal benchmark. All multimodal recordings were de-identified before annotation and analysis, and personally identifiable information was removed from the research dataset. Detailed procedures for session-level distress ratings, tic-frequency assessment, acquisition settings, and annotation rules are provided in Supplementary Methods Sections 6.1–6.2.

MindTS-Net framework

MindTS-Net was designed as a closed-loop multimodal intervention framework linking sensing, emotion perception, strategy planning, guidance generation, and dialogue-state tracking. As shown in Figure 6, speech, transcript, and visual inputs were first captured and temporally aligned, and then encoded into modality-specific representations. These representations were fused to estimate the child’s emotional state and to support downstream strategy selection and guidance generation.

Figure 6.

Architecture of MindTS-Net. T

The framework consisted of four core modules. The Emotion Perception Module (EPM) encoded and fused text, speech, and visual cues into a unified emotional representation. The Mindfulness Intervention Planner (MIP) mapped the emotional representation to one of the predefined mindfulness strategies, such as breathing regulation, body scan, emotion labeling, mindfulness attention, relaxation training, and awareness guidance. The Mindful Text Generator (MTG) generated age-appropriate guidance conditioned on the estimated emotion and selected strategy. The Dialogue State Tracker (DST) maintained multi-turn information about emotional states, selected strategies, generated responses, and tic-related features to support adaptive adjustment across turns.

MindTS-Net was trained using a multi-task objective in which emotion recognition, strategy planning, text generation, and dialogue-state tracking were optimized jointly. The overall objective was formulated as a weighted sum of task-specific losses:

$$\mathcal{L}=\mathit{\alpha }{\mathcal{L}}_{\mathit{e}\mathit{p}\mathit{m}}+\mathit{\beta }{\mathcal{L}}_{\mathit{m}\mathit{i}\mathit{p}}+\mathit{\gamma }{\mathcal{L}}_{\mathit{m}\mathit{t}\mathit{g}}+\mathit{\delta }{\mathcal{L}}_{\mathit{d}\mathit{s}\mathit{t}}$$ (Equation 1)

where ${\mathcal{L}}_{epm},{\mathcal{L}}_{mip},{\mathcal{L}}_{mtg}$ and ${\mathcal{L}}_{dst}$ denote the loss functions for emotion perception, mindfulness intervention planning, mindful text generation, and dialogue-state tracking, respectively, and α, β, γ, and δ denote the corresponding task weights.

Emotion perception module

The Emotion Perception Module (EPM) encoded temporally aligned text, audio, and visual inputs and fused them into a unified emotional representation. This representation integrated semantic, acoustic, and visual affective cues and was used to support downstream strategy planning and mindfulness guidance generation. The detailed architecture of EPM is shown in Figure 7.

Figure 7.

Detailed architecture of EPM

EPM processed three perceptual channels: text x^T from ASR transcripts, audio x^A from speech waveforms, and visual x^V from image frames capturing facial expression and posture. These channels represented semantic, acoustic, and visual dimensions of emotion and were integrated into a unified emotional state representation:

$$\mathit{e}=f_{\text{EPM}}\left(x^T,x^A,x^V\right)\in {\mathbb{R}}^{d_e}$$ (Equation 2)

To address differences in sampling rate and latency, each text segment at timestamp t_i was aligned with the temporally nearest audio and visual frames within a predefined time window Δt, ensuring synchronized multimodal inputs for encoding and fusion.

For modality-specific encoding, text inputs were processed by a pre-trained Chinese-BERT encoder to generate contextualized embeddings:

$$h^T=\text{BERT}\left(x^T\right)\in {\mathbb{R}}^{d_T}$$ (Equation 3)

The audio waveform was transformed into a Mel-spectrogram using short-time Fourier transform and then modeled by stacked convolutional layers followed by a BiGRU:

$$h^A=\text{BiGRU}\left(\text{CNN}\left(\text{Mel}\left(x^A\right)\right)\right)\in {\mathbb{R}}^{d_A}$$ (Equation 4)

For the visual modality, ResNet-50 extracted frame-level visual features, followed by a multilayer perceptron:

$$h^V=\text{MLP}\left(\text{ResNet}50\left(x^V\right)\right)\in {\mathbb{R}}^{d_V}$$ (Equation 5)

COVAREP and OpenFace were used for preprocessing support and quality control, but they were not treated as the primary trainable inputs to the EPM encoders.

Multimodal features were fused using a Cross-Modal Transformer, with text serving as the primary modality and audiovisual streams serving as auxiliary inputs. Let h^T denote the text feature and h^M = [h^A;h^V] denote the concatenated audiovisual representation. The Cross-Modal Transformer applied scaled dot-product attention, where textual features were projected as queries and audiovisual features were projected as keys and values:

$${\tilde{h}}^T=\text{softmax}\left(\frac{Q{K}^{\top }}{\sqrt{d_k}}\right)V$$ (Equation 6)

where Q = W _Qh^T, K = W_Kh^M, V = W_Vh^M. The attended output was combined with the original textual feature through a residual connection and layer normalization to obtain the final emotion vector:

$$\mathrm{e}=\text{LayerNorm}\left({\tilde{h}}^T+h^T\right)\in {\mathbb{R}}^{{\mathrm{d}}_{\mathrm{e}}}$$ (Equation 7)

This fused representation integrated semantic, acoustic, and visual affective information and served as the emotional representation for downstream strategy planning and guidance generation.

Mindfulness intervention planner

The Mindfulness Intervention Planner (MIP) translated the fused emotional representation from EPM into a tailored mindfulness strategy, which then served as the control signal for downstream mindfulness-oriented text generation. The structure of MIP is shown in Figure 8.

Figure 8.

Structure of MIP

To meet the therapeutic needs of children with TS, we defined a fixed set of standardized intervention strategies $\mathcal{S}=\left\{S_1,S_2,\dots ,S_K\right\}$ grounded in clinical practice and established psychological principles. In the present study, MIP included six strategies: breathing regulation (S₁), body scan (S₂), emotion labeling (S₃), mindfulness attention (S₄), relaxation training (S₅), and awareness guidance (S₆).

In the closed-set setting, strategy selection was formulated as a multi-class classification problem. The fused emotional representation e_fuse was passed through a lightweight feedforward classifier with ReLU activation and Softmax normalization to obtain strategy probabilities, and the most probable strategy was selected:

$$\hat{y}=\text{agr}{\max }_{k}p\left(S_k\mid {\mathit{e}}_{\text{fuse}}\right)$$ (Equation 8)

where $\hat{y}$ denotes the predicted strategy label, and K = 6 denotes the number of predefined strategy categories. The selected label was then mapped to a predefined strategy template t_k. The corresponding template embedding was concatenated with the contextual representation e_context to form the decoder control input:

$${\mathit{x}}_{\text{input}}=[\text{Embed}\left(t_k\right);{\mathit{e}}_{\text{context}}]$$ (Equation 9)

This control vector initialized downstream generation and constrained the generated guidance to remain consistent with the selected intervention strategy.

For scenarios requiring dynamic expansion or replacement of strategies, MIP also supported open-set retrieval through semantic matching in the embedding space. Each candidate strategy description t_k was encoded by a BERT-based encoder into a semantic embedding ${\mathit{e}}_{y_k}\in {\mathbb{R}}^d$. The strategy whose embedding was most similar to the fused emotional representation was selected by cosine similarity:

$$\hat{y}=\text{agr}{\max }_k\frac{{\mathit{e}}_{\text{fuse}}·{\mathit{e}}_{y_k}}{\Vert {\mathit{e}}_{\text{fuse}}\Vert \Vert {\mathit{e}}_{y_k}\Vert }$$ (Equation 10)

The selected strategy embedding was then concatenated with the contextual representation and projected into the control space:

$${\mathit{z}}_{\text{input}}=W_z\left[{\mathit{e}}_{\text{context}};{\mathit{e}}_{y_k}\right]+b_z$$ (Equation 11)

This control representation was passed to the text generator to support personalized and contextually relevant mindfulness guidance.

Mindful text generator

The Mindful Text Generator (MTG) produced personalized mindfulness guidance using a prefix-injected Transformer conditioned on the fused emotional representation from EPM and the strategy representation from MIP. The architecture of MTG is shown in Figure 9.

Figure 9.

Architecture of MTG

Let $\mathit{e}\in {\mathbb{R}}^{d_e}$ denote the emotional state vector from EPM and $\mathit{s}\in {\mathbb{R}}^{d_s}$ denote the strategy vector from MIP. These two control signals were mapped by nonlinear functions into layer-wise prefix representations for a decoder with L layers:

$${\mathit{P}}_e=f_e\left(\mathit{e}\right)=\left[{\mathit{p}}_e^{\left(1\right)},\dots ,{\mathit{p}}_e^{\left(L\right)}\right]$$

$${\mathit{P}}_s=f_s\left(\mathit{s}\right)=\left[{\mathit{p}}_s^{\left(1\right)},\dots ,{\mathit{p}}_s^{\left(L\right)}\right]$$ (Equation 12)

where P_e and P_s denote the full emotion-prefix set and strategy-prefix set, respectively, and ${\mathit{p}}_e^{\left(l\right)}$ and ${\mathit{p}}_s^{\left(l\right)}$ denote the prefix representations injected into the l-th decoder layer for emotion control and strategy control.

For each decoder layer l ∈ {1, …,L}, let $K_x^{\left(l\right)}$ and $V_x^{\left(l\right)}$ denote the standard decoder key and value states computed from the input token representations at layer l. The corresponding emotion and strategy prefixes were prepended to these states:

$${\mathit{K}}^{\left(l\right)}=\left[{\mathit{p}}_e^{\left(l\right)};{\mathit{p}}_s^{\left(l\right)};K_x^{\left(l\right)}\right]$$

$${\mathit{V}}^{\left(l\right)}=\left[{\mathit{p}}_e^{\left(l\right)};{\mathit{p}}_s^{\left(l\right)};V_x^{\left(l\right)}\right]$$ (Equation 13)

The resulting decoder attention followed the standard scaled dot-product form, allowing the generated response to be jointly conditioned on both the emotional state and the selected intervention strategy.

To ensure age-appropriate expression, MTG applied a cognitive normalization step:

$$Y^{\prime }=\text{Normalize}\left(\hat{Y}\right)$$ (Equation 14)

This step used a mindfulness terminology dictionary, simplified syntax rules, and emotion-sensitive templates to produce developmentally appropriate wording and a mindfulness-consistent tone.

Instead of repeatedly regenerating responses within the same turn, the model evaluated response effectiveness using post-turn emotional feedback from the subsequent interaction. The observed post-turn emotional state e′ was compared with the prior state e, and the magnitude of emotional change was computed as:

$${\delta }_{\text{emo}}={\Vert {\mathit{e}}^{\prime }-\mathit{e}\Vert }_2$$ (Equation 15)

This feedback signal was passed to the Dialogue State Tracker to guide strategy adaptation in later turns. To reduce unsafe or inappropriate outputs, generated responses were additionally constrained by rule-based safety filtering and researcher review procedures for high-risk interaction content.

Dialogue state tracker

The Dialogue State Tracker (DST) maintained and updated a multi-turn state representation integrating emotional representations, selected strategies, generated responses, and tic-related signals. This module supported dynamic adaptation across dialogue turns by retaining recent interaction history and updating strategy selection according to emotional and tic-related changes. The structure of DST is shown in Figure 10.

Figure 10.

Structure of the DST module

At dialogue turn t, DST received a composite state tuple:

$$S_t=\left\{{\mathit{e}}_t,{\mathit{s}}_t,{\mathit{y}}_t,{\mathit{h}}_t^{\text{tic}}\right\}$$ (Equation 16)

where e_t denotes the emotional representation from EPM, s_t denotes the selected intervention strategy from MIP, y_t denotes the guidance representation generated by MTG, and ${\mathit{h}}_t^{\text{tic}}$ denotes the tic-related state. DST updated the dialogue context recursively using a GRU:

$${\mathit{h}}_t^{\text{dialog}}=\text{GRU}\left(S_t,{\mathit{h}}_{t-1}^{\text{dialog}}\right)$$ (Equation 17)

The hidden state ${\mathit{h}}_t^{\text{dialog}}$ summarized the dialogue history up to turn t and conditioned subsequent response generation. To reduce redundancy, DST retained only the most recent K turns and applied an attention-based weighting scheme over this fixed-length memory window, allowing more informative recent states to receive greater emphasis before recurrent updating.

The effectiveness of the current intervention was evaluated jointly from emotional change and tic-state change. Specifically, DST computed an intervention effectiveness score:

$${\gamma }_t={\lambda }_1{\Vert \Delta {\mathit{e}}_t\Vert }_2+{\lambda }_2{\Vert \Delta {\mathit{h}}_t^{\text{tic}}\Vert }_2$$ (Equation 18)

where Δe_t and $\Delta {\mathit{h}}_t^{\text{tic}}$ denote turn-level changes in emotional state and tic-related state, respectively, and λ₁ and λ₂ control their relative contributions. If γ_t exceeded the historical mean, the current strategy was reinforced by increasing its reuse probability in similar states. Conversely, if γ_t remained below a predefined threshold ε for M consecutive turns, the current strategy was treated as insufficiently effective, and MIP updated or switched the strategy in subsequent interactions.

To capture longitudinal regulation patterns, DST smoothed the dialogue-state trajectory using a sliding window:

$${\overline{\mathit{h}}}_t^{\left(w\right)}=\frac{1}{w}\sum _{i=t-w+1}^t{\mathit{h}}_i^{\text{dialog}}$$ (Equation 19)

Principal component analysis was then applied to the smoothed trajectory to extract low-dimensional emotion–behavior trends. These trend representations summarized longer-term changes in emotion regulation, tic dynamics, and intervention latency, and were used to update the user profile vector u and refine the personalized strategy space ${\mathcal{S}}_u\subset \mathcal{S}$.

Quantification and statistical analysis

Dataset-level model performance was evaluated using task-specific metrics. Discrete emotion recognition was assessed using Accuracy, Precision, Recall, and Macro-F1, while continuous Valence–Arousal estimation was assessed using Pearson’s correlation coefficient (PCC) and mean squared error (MSE). Mindfulness strategy selection was evaluated using Strategy Accuracy and consistency metrics, and guidance generation quality was evaluated using BLEU-1, BLEU-4, ROUGE-L, BERTScore, cosine similarity, Distinct-1, Distinct-2, and expert rating scores. Unless otherwise specified, dataset-level metrics were computed on the held-out test set.

Clinical analyses were conducted at the participant level. Repeated daily observations were aggregated within participants before inferential testing. Emotional outcomes included changes in Valence and Arousal between baseline and intervention phases, while tic-related outcomes included observed daily tic counts and tic frequency reduction rate (TFRR). Between-group comparisons were performed using participant-level change scores. Strategy-level analyses summarized emotional and tic-related changes according to the dominant mindfulness strategy assigned during intervention sessions. Because strategies were adaptively assigned and participants could contribute to multiple strategy categories, these analyses were interpreted as exploratory and non-causal rather than as evidence of differential therapeutic efficacy.

External validation analyses compared the primary MindTS-Net cohort with the independent external validation cohort. The external cohort was not used for model training, validation, hyperparameter tuning, or model selection. Cross-center analyses included emotion perception performance, strategy-selection consistency, guidance-generation quality, intervention completion rate, Valence–Arousal changes, observed daily tic counts, and TFRR. Because the external cohort did not include a separate control group, external validation results were interpreted as evidence of cross-center consistency rather than controlled external efficacy. Continuous variables were summarized as mean ± standard deviation unless otherwise specified. Effect sizes were reported using Cohen’s d with 95% confidence intervals where applicable. Model training used participant-level data splitting to reduce identity leakage, random seeds were fixed across repeated experiments whenever applicable, and early stopping was applied according to validation loss. Statistical analyses were performed using Python and R. Unless otherwise specified, all tests were two-sided and statistical significance was defined as p < 0.05.