Nurhaliza, Mutiara (2026) Deteksi Dysgraphia (Specific Learning Disorder with Impairment in Written Expression) pada Anak Menggunakan CNN dan Explainable AI dengan Sistem Feedback Otomatis. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 5027221010-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
5027221010-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy |
Abstract
Dysgraphia merupakan gangguan belajar pada aspek ekspresi tulisan yang kerap terlambat teridentifikasi karena skrining di lapangan masih dominan mengandalkan penilaian manual yang memerlukan waktu, konsistensi antar-evaluator, serta akses terhadap tenaga ahli yang tidak selalu tersedia merata. Keterlambatan identifikasi dapat menghambat intervensi dini dan perencanaan dukungan belajar yang tepat. Penelitian ini mengembangkan sistem skrining dini dysgraphia berbasis citra tulisan tangan melalui aplikasi Android dengan pendekatan Convolutional Neural Network (CNN) yang dipadukan dengan Explainable AI (Grad-CAM) serta feedback otomatis berbasis aturan, sehingga keluaran sistem tidak hanya berupa hasil klasifikasi, tetapi juga penjelasan visual dan rekomendasi perbaikan yang mudah dipahami. Data penelitian mencakup 345 citra tulisan tangan dari sumber eksternal, sintetis, dan riil dengan dua kategori, yaitu potential dysgraphia (PD) dan low potential dysgraphia (LPD). Evaluasi dilakukan menggunakan Stratified 5-Fold Cross-Validation dengan metrik loss, akurasi, presisi, recall, dan AUC, serta pembandingan terhadap beberapa model transfer learning. Hasil menunjukkan CNN Model usulan memberikan performa terbaik pada validasi dengan loss 0,2889; akurasi 92,78%; presisi 93,33%; recall 91,33%; dan AUC 0,9419. Mekanisme feedback diturunkan dari ambang berbasis Youden Index agar rekomendasi lebih spesifik, sementara Grad-CAM menampilkan heatmap area tulisan yang paling berkontribusi terhadap keputusan model untuk meningkatkan transparansi interpretasi. Implementasi pada aplikasi Android berjalan stabil dan responsif, mendukung skrining awal yang lebih praktis
=====================================================================================================================================
Dysgraphia is a learning disorder affecting written expression that is often identified late because screening in practice still relies heavily on manual assessment, which is time-consuming, requires consistent inter-rater judgment, and depends on specialist availability that is not evenly distributed. Delayed identification can hinder early intervention and the planning of appropriate learning support. This study develops an early dysgraphia screening system using handwriting images on an Android application, employing a Convolutional Neural Network (CNN) combined with Explainable AI (Grad-CAM) and rule-based automated feedback, so the system outputs not only a classification result but also visual explanations and understandable improvement recommendations. The dataset comprises 345 handwriting images from external, synthetic, and real sources in two categories: potential dysgraphia (PD) and low potential dysgraphia (LPD). Model performance was evaluated using Stratified 5-Fold Cross-Validation with loss, accuracy, precision, recall, and AUC metrics, alongside comparisons with several transfer learning models. Results show that the proposed CNN Model achieved the best validation performance with loss 0.2889, 92.78% accuracy, 93.33% precision, 91.33% recall, and 0.9419 AUC. The automated feedback mechanism was derived using a Youden Index–based threshold to provide more specific recommendations, while Grad-CAM produced heatmaps highlighting handwriting regions that contributed most to the model’s decision, improving interpretability. The Android implementation ran stably and responsively, supporting a more practical early screening process.
Actions (login required)
 |
View Item |