Rancang Bangun Sistem Deep Learning untuk Klasifikasi Tingkat Depresi Melalui Analisis Wajah dan Respons Kuesioner Berbasis Website

Abstract

Depresi pada mahasiswa menjadi persoalan akademik dan kesehatan publik karena dapat menurunkan konsentrasi, motivasi belajar, capaian akademik, serta meningkatkan risiko isolasi sosial dan ide bunuh diri. Penelitian ini bertujuan merancang dan mengevaluasi sistem skrining awal tingkat depresi berbasis website yang mengintegrasikan kuesioner Patient Health Questionnaire-9 (PHQ-9) dan analisis ekspresi wajah berbasis Convolutional Neural Network (CNN). Metode pengembangan menggunakan model Waterfall yang mencakup komunikasi kebutuhan, perencanaan arsitektur, pemodelan, konstruksi, pengujian, dan deployment. Model visual dilatih menggunakan dataset ekspresi wajah FER-2013 dan CK+ yang diproses menjadi citra grayscale 48x48 piksel, diseimbangkan dengan oversampling, serta diperkuat melalui augmentasi. Sistem menerapkan decision-level fusion dengan bobot 60% untuk skor PHQ-9 dan 40% untuk rasio ekspresi negatif, sedangkan inferensi wajah dijalankan langsung di peramban melalui TensorFlow.js untuk menjaga privasi biometrik. Hasil pengujian menunjukkan model Custom CNN Mini-VGG memperoleh akurasi global 65,28% pada 7.178 citra uji, dengan kinerja terbaik pada kelas Happy dan Surprise. Pengujian fungsional membuktikan autentikasi, validasi webcam, fusi data, ekspor laporan, dan dasbor institusi berjalan valid. Sistem juga terbukti tidak mengirim citra wajah ke server, sehingga prinsip Zero Data Retention terpenuhi. Temuan ini menunjukkan bahwa pendekatan multimodal dapat memperkuat skrining dini depresi, tetapi hasilnya tetap harus diposisikan sebagai triase awal, bukan diagnosis klinis.

Depression among university students is an academic and public health concern because it may reduce concentration, learning motivation, academic achievement, and increase the risk of social withdrawal and suicidal ideation. This study aims to design and evaluate a web-based early screening system for depression severity by integrating the Patient Health Questionnaire-9 (PHQ-9) and facial expression analysis using a Convolutional Neural Network (CNN). The system was developed using the Waterfall model, covering requirement communication, architectural planning, modeling, construction, testing, and deployment. The visual model was trained on FER-2013 and CK+ facial expression datasets, standardized into 48x48 grayscale images, balanced using oversampling, and strengthened with data augmentation. The proposed system applies decision-level fusion, assigning 60% weight to PHQ-9 scores and 40% to the negative facial expression ratio, while facial inference is executed locally in the browser through TensorFlow.js to protect biometric privacy. The evaluation shows that the Custom CNN Mini-VGG model achieved 65.28% global accuracy on 7,178 unseen test images, with the strongest performance for Happy and Surprise classes. Functional testing confirmed that authentication, webcam validation, data fusion, report export, and institutional dashboards worked as expected. Network inspection also confirmed that no facial images were transmitted to the server, thereby satisfying the Zero Data Retention principle. These findings indicate that multimodal screening can strengthen early depression triage, although the system should remain a decision-support tool and not a substitute for clinical diagnosis.