Early screening for dyslexia & specific learning difficulties

Reading and spelling difficulties are often noticed only after a child has failed for years. Short AI-supported screenings (eye tracking while reading, gamified tasks, spoken phonological tasks) estimate the risk of dyslexia or related difficulties in minutes, so school psychologists and special-education teachers can prioritise children for full assessment and start support early. The output is decision support, never a diagnosis.

Problem & value

• Today: screening relies on teacher observation, paper-based tests (e.g. phonological awareness checklists, spelling tests) and long clinical assessments with waiting lists. Children with difficulties are identified late, often after emotional and academic consequences.
• AI value: standardised, quick, repeatable screening of whole classes; objective signals (eye movements, response timing) that teachers cannot observe; triage so scarce specialists see the right children first.
• Evidence: published research shows that eye-movement features recorded during reading can separate dyslexic from typical readers in study samples [1], and a gamified online test with machine learning has been evaluated for risk prediction [2]. Results come from specific samples and languages; real-world accuracy in a different population must be validated locally, and no universal figures should be assumed.

Variants

• Eye-tracking reading screen: the child reads a short text on a screen; reading behaviour is analysed (Lexplore, based on Karolinska Institutet research, used in Swedish and UK schools) [3].
• Game-based screener: tablet or web games measuring language and perceptual skills, classified by ML (Dytective by Change Dyslexia, Spain) [4].
• Spoken phonological tasks: speech recognition scores rhyming, syllable segmentation, non-word repetition and rapid naming; overlaps with EDU-012 (reading fluency).
• Kindergarten / pre-school risk screening: before formal reading instruction, using early literacy precursors.
• Dyscalculia and attention-related difficulties: same pipeline with other task batteries.
• Clinical-diagnostic version: a tool that claims to diagnose or treat is likely a medical device under the MDR [5] and moves to the health domain (see HLTH-001).
• Private use: parents using consumer apps; same limits, no school duty of care.

How it works

1. consent & set-up (guardian information, language, device calibration)
2. task delivery (reading on eye tracker | game on tablet | spoken prompts)
3. signal capture (gaze, taps/latencies, audio)
4. quality checks (calibration, noise, attention drop-outs)
5. feature extraction (fixations, regressions, error patterns, phoneme-level scores)
6. risk model (calibrated classifier, age/language-normed)
7. report with uncertainty and explanation
8. human review (school psychologist / special-education teacher)
9. referral for full assessment or support plan; parents informed

• Models: mostly classical ML (gradient boosting, SVM, small neural nets) on engineered features and time series; ASR for speech tasks. LLMs are only used to draft plain-language reports.
• Norming: scores must be compared with age- and language-specific norms; bilingual children and dialect speakers need separate checks.
• Human-in-the-loop: every elevated-risk result is reviewed by a professional; the diagnosis (e.g. under ICD-11 or national guidelines) stays with qualified clinicians.

Data preparation

Unlike document-based use cases, the data here is mostly generated fresh by the screening itself: gaze recordings, tap and latency logs, and audio of spoken tasks. What exists already sits in school management software, paper files of the school psychologist, teacher observation sheets and earlier test results (often Excel or paper), plus class lists with home language and bilingual status.

Typical gaps: no expert-confirmed outcome labels (the key asset for any own model), scattered and unstructured teacher notes, inconsistent language and dialect information, poor calibration or noisy audio, missing guardian consents, and unclear access rights to sensitive learning-difficulty data.

Preparation steps: - Define the data schema and the outcome to be predicted (see PREP-019) and harvest existing assessment results as labels where lawful (PREP-020). - Digitise paper test results and observation sheets (PREP-006, PREP-008) and tidy spreadsheets (PREP-011). - Build a local evaluation set with school psychologists (PREP-018). - Pseudonymise exports (PREP-022), check the legal basis (PREP-025) and restrict access strictly (PREP-024).

Effort share: medium for buyers (consents, device set-up, class data import), high for own models, where labelling and follow-up data dominate. See also GUIDE-003 for a readiness check.

Implementation paths

Which path for whom

• Enterprise (e.g. school operators, education providers): Path A, because validated screeners with language norms exist and a self-built risk model would need clinical validation first.
• Public (schools, school authorities): Path A with an EU-hosted, validated product, because procurement, duty of care and sensitive children's data favour a certified, supported tool.
• Developer (vendors, research groups): Path C, because the value lies in owning the feature models and on-device scoring (Whisper, scikit-learn on tablets), so raw child data stays local.

Path A — Buy (recommended)

• Eye tracking: Lexplore (Sweden, EU provider) [3].
• Game-based: Dytective / Change Dyslexia (Spain, EU provider) [4].
• Broader literacy platforms: national or regional screening instruments from education ministries and publishers; check whether the tool is validated for your language.
• Choose A when: a school authority or school needs a validated, supported tool. Ask for validation studies in your language, sensitivity/specificity with confidence intervals, norm data, EU hosting and a data processing agreement.

Path B — Build on cloud APIs

• Cloud ASR in EU regions (Azure AI Speech, Google Cloud Speech-to-Text, Mistral Voxtral) for phonological tasks; LLM only for report wording, with pseudonymised data.
• Child speech is hard for generic ASR; use pronunciation-assessment features and test error rates on children of the target age before relying on it.
• Rarely sufficient alone: the risk model needs clinically labelled data.

Path C — Self-host (open source)

• Speech: Whisper (OpenAI, open weights) [6], wav2vec 2.0 / HuBERT-style models fine-tuned on child speech; forced alignment for phoneme-level scoring.
• Gaze: consumer webcams give weak accuracy; dedicated eye trackers (e.g. Tobii, EU) or open libraries such as WebGazer are research grade only.
• ML: scikit-learn / XGBoost, scored on-device (tablet) so raw child data stays local.
• Hardware: HW1 (CPU / tablet) is sufficient for feature models; HW2 only for larger ASR.

Path D — Custom training

Worthwhile for a research group, ministry or vendor with a new language. Needs hundreds to thousands of children with expert-confirmed outcomes, longitudinal follow-up (screen early, confirm later), and ethics approval. Plan stratified validation across languages, dialects, gender, socio-economic background and devices; report calibration, not just accuracy.

State of the art (2026)

• Eye tracking and game-based screeners are established in some national school systems; peer-reviewed studies report good separation of groups in research samples [1][2], but sample sizes, languages and designs differ, and published accuracy is often higher than prospective school performance.
• Trend: multimodal signals (gaze + speech + handwriting/typing), longitudinal screening instead of one-off tests, and on-device processing for privacy.
• Limits and failure modes: false positives cause stigma and anxiety; false negatives give false reassurance; poor norming for multilingual children, dialects and regular-orthography languages; device and calibration artefacts; fatigue or attention problems mimic risk; comorbidities (ADHD, hearing or vision problems) confound results; automation bias in staff.

Shared resources & combinations

• Shared stack: ASR and pronunciation scoring with EDU-012 and EDU-011; tabular-ml and time-series with EDU-007; on-device runtime with PERS-019.
• Downstream: EDU-005 adaptive learning paths and EDU-009 teacher differentiation can use the profile (with consent) to personalise training; EDU-003 for feedback on spelling tasks.
• Boundary: HLTH-001 for clinical diagnosis tools under MDR.

Effort, cost & KPIs

• Path A: weeks to a school term; licences usually per pupil or per screening, plus devices (eye tracker or tablets) and teacher training.
• Path B/C: 6-12 months including validation; main cost is data collection and expert labelling.
• Path D: quarters to years, research-project scale.
• KPIs: sensitivity and specificity against clinical reference, positive predictive value in your base rate, calibration, share of children screened, time from flag to support, subgroup parity (language, gender, background), later reading progress, parent and teacher acceptance.

People & change

Involve management and the school authority first (decision, budget, no screening without a support offer), then the school psychologists and special-education teachers who review every flag, and IT for devices and accounts. Bring the data protection officer in before any pilot: children's gaze and voice data and inferred learning-disability status require a DPIA, a legal basis and an AVV with the vendor (GUIDE-005). Prepare guardian information and a clear appeal route; parents are the affected stakeholders.

Works council: depends. The tool evaluates pupils, not employees, so §87(1) no. 6 BetrVG is usually not triggered. But teachers' usage logs and handling can be monitored, so it applies if the tool records staff activity in a company-run school. Inform the works council early under §90(1) no. 3 BetrVG; public schools involve the staff council (Personalrat) under BPersVG or state law (GUIDE-004).

Support AI literacy of teachers (AI Act Art. 4): what the score means, false positives and negatives, automation bias (GUIDE-006). Choose one grade and a few motivated teachers as pilot group, compare flags with their judgement, and communicate openly that the result is decision support, never a diagnosis (GUIDE-008). Typical concerns are stigma, extra workload and labelling; address them with professional review, short retention and support planning attached to every flag.

Risks, regulation & compliance

• EU AI Act [7]: Annex III point 3 covers AI used to determine access or admission, assign persons to educational institutions or programmes, evaluate learning outcomes, assess the appropriate level of education, and monitor prohibited behaviour in tests. A screener used only for support planning may fall outside, but if its result decides placement (e.g. special-needs school or track), treat it as high risk: risk management, data governance, bias testing, human oversight, logging, transparency to deployers, and deployer duties (Art. 26, incl. informing affected persons). Hence depends.
• Art. 5 prohibition: emotion inference in educational institutions is prohibited (Art. 5(1)(f)). Do not infer emotions, attention or stress from gaze or facial data; restrict analysis to reading skills. Biometric categorisation of sensitive traits is also prohibited.
• MDR [5]: claims of diagnosis or medical purpose can make the software a medical device. Word intended use carefully ("screening support, no diagnosis") and get regulatory advice.
• GDPR [8]: children's data, gaze and voice recordings, and inferred learning-disability status (health-related, Art. 9) need a clear legal basis (public task or explicit guardian consent), DPIA, data minimisation, short retention, EU hosting, no use of children's data for vendor model training without separate basis, and information to parents. Art. 22 limits solely automated decisions.
• Sector rules: national school and special-education law (e.g. German Laender rules on Nachteilsausgleich and Foerderdiagnostik) decide who may screen and what follows.
• Mitigations: professional review, documented limits, local validation, subgroup monitoring, appeal route for parents, deletion of raw recordings.

Getting started

1. Define the decision: who is flagged, who reviews, what support follows (no screening without support).
2. Check candidate tools for validation in your language, norms, MDR status, and AI Act positioning.
3. Run a DPIA and prepare guardian information and consent materials.
4. Pilot with one grade, compare flags with teacher judgement and later professional assessment.
5. Monitor subgroup error rates and reading progress; adjust thresholds with the school psychologist.

Sources

1. Benfatto et al. 2016 — Screening for dyslexia using eye tracking during reading (PLOS ONE) — https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0165508
2. Rello et al. 2020 — Predicting risk of dyslexia with an online gamified test (PLOS ONE) — https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0241687
3. Lexplore — eye-tracking reading assessment — https://www.lexplore.com
4. Change Dyslexia / Dytective — https://dytective.com
5. Regulation (EU) 2017/745 (MDR) — https://eur-lex.europa.eu/eli/reg/2017/745/oj
6. OpenAI Whisper — https://github.com/openai/whisper
7. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
8. Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj