Electronic Theses and Dissertation

Disleksia merupakan gangguan belajar neurologis yang memengaruhi kemampuan
membaca dan menulis, dengan prevalensi diperkirakan mencapai 10–20% pada anak di
Indonesia. Diagnosis konvensional umumnya masih bergantung pada asesmen
psikometrik yang bersifat subjektif serta memerlukan waktu yang relatif lama.
Penelitian ini mengusulkan pendekatan Explainable Machine Learning berbasis
algoritma Random Forest dan metode interpretabilitas SHAP (SHapley Additive
exPlanations) untuk menganalisis pola eye-tracking pada anak disleksia dan non
disleksia menggunakan dataset ETDD70. Dataset tersebut memuat data eye-tracking
dari 70 anak berusia 9–10 tahun, yang terdiri atas 35 anak disleksia dan 35 anak non
disleksia, yang menjalani tiga tugas membaca terkontrol. Sebanyak 24 fitur akhir (8 fitur
× 3 tugas) diperoleh melalui proses seleksi fitur berjenjang berdasarkan analisis korelasi,
pemeriksaan variansi, dan justifikasi ilmiah. Model dilatih dan dievaluasi menggunakan
skema Stratified 5-Fold Cross-Validation pada level subjek untuk mencegah kebocoran
data (data leakage). Hasil evaluasi menunjukkan performa model yang baik dengan
accuracy 0,857, precision 0,906, recall 0,800, F1-score 0,843, dan AUC-ROC 0,906.
Analisis SHAP mengungkapkan bahwa fitur dwell_time_trial pada tugas T1, T4,
dan T5 merupakan kontributor utama terhadap prediksi model, dengan
T1_dwell_time_trial menempati peringkat tertinggi secara global, diikuti oleh
n_fix_trial pada tugas T4. Arah kontribusi seluruh fitur konsisten dengan teori
kognitif disleksia, di mana anak disleksia menunjukkan durasi membaca, jumlah fiksasi,
dan regresi yang lebih tinggi, serta amplitudo sakade yang lebih rendah dibandingkan
anak non-disleksia. Penelitian ini menunjukkan bahwa integrasi Random Forest dan
SHAP mampu menghasilkan model klasifikasi disleksia yang tidak hanya akurat, tetapi
juga interpretatif secara ilmiah.
Kata kunci: disleksia, eye-tracking, Random Forest, Explainable Machine Learning,
SHAP, ETDD70

Dyslexia is a neurological learning disorder that affects reading and writing abilities, with an estimated prevalence of 10–20% among children in Indonesia. Conventional diagnosis generally still relies on psychometric assessments that are subjective and require a relatively long time. This study proposes an Explainable Machine Learning approach based on the Random Forest algorithm and the SHAP (SHapley Additive exPlanations) interpretability method to analyze eye-tracking patterns in dyslexic and non-dyslexic children using the ETDD70 dataset. The dataset contains eye tracking data from 70 children aged 9–10 years, consisting of 35 dyslexic children and 35 non-dyslexic children, who underwent three controlled reading tasks. A total of 24 final features (8 features × 3 tasks) were obtained through a tiered feature selection process based on correlation analysis, variance inspection, and scientific justification. The model was trained and evaluated using a subject-level Stratified 5 Fold Cross-Validation scheme to prevent data leakage. The evaluation results show good model performance, with an average accuracy of 0.857, precision of 0.906, recall of 0.800, F1-score of 0.843, and AUC-ROC of 0.906. SHAP analysis reveals that the dwell_time_trial feature in tasks T1, T4, and T5 is the main contributor to the model's predictions, with T1_dwell_time_trial ranking highest globally, followed by n_fix_trial in task T4. The direction of contribution of all features is consistent with cognitive theories of dyslexia, in which dyslexic children show longer reading duration, higher fixation counts, and more regressions, as well as lower saccade amplitude compared to non-dyslexic children. This study demonstrates that the integration of Random Forest and SHAP can produce a dyslexia classification model that is not only accurate but also scientifically interpretable. Keywords: dyslexia, eye-tracking, Random Forest, Explainable Machine Learning, SHAP, ETDD70