Mobile-Based AI Screening Systems for ADHD and Speech Delay: A Survey Based on Recent Clinical, Sensor, and Machine Learning Studies

Abstract

Early and accurate screening for neurodevelopmental disorders, such as attention deficit hyperactivity disorder (ADHD) and speech delay, is crucial for maximizing the effectiveness of early intervention strategies. Traditional diagnostic methods, which rely heavily on expert-administered subjective clinical interviews and standardized questionnaires, are well-validated but often suffer from limitations in accessibility, resource demands, and time consumption, especially in settings with limited resources. The increasing availability of mobile technology, coupled with breakthroughs in artificial intelligence (AI) and multimodal sensing, presents a new opportunity for creating scalable, objective, and readily accessible screening tools. This survey provides a comprehensive review of existing mobile-based AI screening systems. It specifically examines the various data types used (including clinical scales, speech patterns, and visual attention metrics), the machine learning models employed, and the incorporation of explainable AI (XAI) for transparency. Furthermore, the review addresses practical aspects of real- world deployment, including system architectures, ethical concerns related to child data privacy, and promising future research avenues, such as federated and gamified screening approaches.

Introduction

Neurodevelopmental disorders such as ADHD and speech delay are increasingly prevalent worldwide, significantly affecting children’s academic performance, emotional regulation, and social development.

• ADHD is characterized by inattention, hyperactivity, and impulsivity.

• Speech delay involves lagging expressive or receptive language development.

Traditional diagnosis relies heavily on specialists and subjective rating scales, leading to delays, high costs, and limited access—especially in low- and middle-income regions. As a result, many children miss the critical early neuroplastic intervention window.

A major shift toward AI-powered mobile health solutions is emerging to decentralize screening and improve accessibility.

Core Problem

The fundamental issue is a gap between clinical demand and diagnostic capacity.

Key Systemic Challenges:

1. Specialist Shortages → Long wait times and delayed intervention.

2. Subjectivity in Screening → Bias in parent/teacher rating scales.

3. Technological Disparity → Advanced tools (e.g., fMRI) are impractical; most apps lack clinical rigor.

4. Lack of Multimodal Integration → ADHD and speech delay often assessed separately despite high comorbidity.

The need is for objective, accessible, and multimodal screening systems deployable in community settings.

Clinical Background

ADHD

• Classified (DSM-5) into: Inattentive, Hyperactive-Impulsive, Combined types.

• Common tools: Conners’ Rating Scale, Vanderbilt Assessment Scale.

Speech Delay

• Features: limited vocabulary, articulation issues, weak sentence structure.

• Frequently comorbid with ADHD.

• Integrated assessment improves diagnostic value.

Data Acquisition Modalities for Mobile AI Screening

Mobile devices collect multimodal biomarkers across three channels:

1. Acoustic & Linguistic Biomarkers

• Low-level features: Pitch (F0), jitter, shimmer, MFCCs, pause analysis.

• High-level features: Lexical diversity (TTR), syntactic complexity (MLU), sentiment.

2. Visual & Oculometric Biomarkers

• Fixation duration

• Saccadic velocity

• Microsaccades

• Blink rate

3. Kinematic & Interaction Biomarkers

• Actigraphy (movement via accelerometer/gyroscope)

• Touch latency and reaction time variability

Feature Engineering

Before modeling:

• Signal Denoising (e.g., spectral subtraction, video stabilization)

• Dimensionality Reduction (PCA, Recursive Feature Elimination)

AI Architectures

Traditional Machine Learning

Best for small datasets:

• Support Vector Machines (SVM)

• Random Forests

• Gradient Boosting (XGBoost, LightGBM)

Deep Learning

For raw audio/video:

• CNNs → Spectrogram & facial pattern analysis

• RNNs / LSTMs → Temporal behavioral patterns

• Transformers & Attention → Focus on distraction moments

• Pretrained models (BERT, Wav2Vec) via transfer learning

Multimodal Fusion Strategies

Early Fusion

• Combines features before classification

• Captures cross-modal correlations

• High dimensionality, strict synchronization required

Late Fusion

• Separate models per modality

• Combine prediction outputs

• More flexible but less interaction modeling

Recent studies (2024–2025) show multimodal fusion >90% accuracy, outperforming single-modality systems.

Deployment Considerations

Mobile Edge Computing

• Model quantization (32-bit → 8-bit)

• TensorFlow Lite (Android), Core ML (iOS)

Gamification

• Serious games (e.g., reaction-time tapping tasks)

• Narrative-driven speech elicitation

Privacy & Ethics

Federated Learning

• Trains models locally on-device

• Only model updates shared

• Raw child data never leaves device

Explainable AI (XAI)

• SHAP for feature contribution analysis

• Heatmaps to show attention focus areas

• Essential for clinical trust

Research Trends (2020–2025)

• 2020–2021: Digitized questionnaires (high but subjective accuracy)

• 2022–2023: Single-sensor methods (more objective, moderate accuracy)

• 2024–2025: Multimodal fusion (high accuracy + robustness)

Age-specific findings:

• Speech markers → More predictive (ages 2–4)

• Attention/gaze markers → More predictive (ages 5–8)

Open Challenges

1. Data Scarcity – Lack of large-scale, open NDD datasets.

2. Longitudinal Stability – Need continuous monitoring instead of single snapshots.

3. Cultural Bias – Models trained on Western populations may not generalize globally.

Conclusion

The convergence of mobile sensing and artificial intelligence represents a watershed moment in developmental pediatrics. This survey has delineated the technical roadmap for building effective ADHD and speech delay screening systems. We established that multimodal fusion—integrating acoustic, visual, and kinematic data is superior to unimodal approaches. We highlighted the necessity of Deep Learning for feature extraction but emphasized the role of Explainable AI for clinical adoption. Finally, we argued that the future lies in privacy-preserving Federated Learning architectures that can scale globally while respecting individual rights. The engineering tools exist; the challenge now is clinical validation and ethical deployment to ensure no child is left behind.

References

The following is a list of references focusing primarily on the application of machine learning and technology for the detection and screening of Attention-Deficit/Hyperactivity Disorder (ADHD) and other neurodevelopmental disorders. [1] Machine Learning for ADHD Detection (Various Modalities) • Interpretable ML for Child ADHD (Clinical Data): H. Qin, Y. Li, and J. Wang, “Interpretable machine learning approaches for children’s ADHD detection using clinical assessment data,” BMC Psychiatry, vol. 25, no. 1, p. 12, 2025. • Dual-Stream ML for Adult ADHD (Symptoms): C. Nash, R. Nair, and S. M. Naqvi, “Insights into detecting adult ADHD symptoms through advanced dual-stream machine learning,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 32, pp. 3378–3389, 2024. [2] ML from FNIRS Signal (Reverse Stroop Tasks) • M. Maniruzzaman, A. Asaduzzaman, and M. Shin, “Machine learning-based ADHD detection from fNIRs signal during reverse stroop tasks,” IEEE Access, vol. 12, pp. 82984–82995, 2024. • Accurate Identification using ML: N. Alsharif, M. H. Al-Adhaileh, and M. Al-Yaari, “Accurate identification of Attention-Deficit/Hyperactivity Disorder using machine learning approaches,” Journal of Disability Research, vol. 3, no. 1, pp. 45–58, 2024. • Entropy Difference-Based EEG Channel Selection: S. Maheshwari, P. Kumar, and R. Gupta, “Novel entropy difference-based EEG channel selection technique for automated detection of ADHD,” IEEE Transactions on Instrumentation and Measurement, vol. 73, pp. 1–9, 2024. • Early Detection and Treatment (Sri Lankan Children): R. S. Deshpande and S. S. Thorat, “Early detection and effective treatment for ADHD using machine learning for Sri Lankan children,” in 2023 5th International Conference on Advancements in Computing (ICAC), Colombo, Sri Lanka, 2023, pp. 345–350. [3] Technology for Screening and Diagnosis (Eye Tracking & Speech) • Innovative Portable Eye Tracking for Screening: J. H. Yoo, C. S. Kang, and J. S. Lim, “Development of an innovative approach using portable eye tracking to assist ADHD screening: A machine learning study,” Frontiers in Psychiatry, vol. 14, p. 125602, 2024. • AI-Based Eye Tracking for Child ADHD: Y. Tian, J. Ye, and H. Liu, “Utilizing artificial intelligence- based eye tracking technology for screening ADHD symptoms in children,” Frontiers in Psychiatry, vol. 14, p. 109873, 2023. • Review on Sensors for Visual Attention: X. Li, Y. Zhang, and M. Chen, “Sensors for human visual attention behavior detection and prediction: A review,” IEEE Sensors Journal, vol. 25, no. 3, pp. 2021–2035, 2025. • Mobile-Based Speech Analysis Review (Developmental Disorders): S. P. Porfiri, M. A. P. Facchinetti, and A. L. V. Cernadas, “Mobile-based speech analysis for early detection of developmental disorders: A systematic review,” IEEE Transactions on Biomedical Engineering, vol. 69, no. 8, pp. 2450–2462, 2022 • Speech Delay Assistive Device (ML for Transcription): P. M. Silva, F. Oliveira, and T. Sousa, “Speech delay assistive device for speech-to- text transcription using machine learning,” Computers, vol. 13, no. 1, p. 60, 2024.

Copyright

Copyright © 2026 Harish R, Karthick Raja V, Parthiban V, A. Sivaramakrishnan, R. Anitha. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.