Evaluasi Performa Klasifikasi Depresi Berbasis Teks Menggunakan TF-IDF dan Transformer Pretrained (BERT, RoBERTa, DistilBERT, XLM-R) pada Dataset Wawancara Klinis DAIC-WOZ

Alwin, Muhammad Ariq (2026) Evaluasi Performa Klasifikasi Depresi Berbasis Teks Menggunakan TF-IDF dan Transformer Pretrained (BERT, RoBERTa, DistilBERT, XLM-R) pada Dataset Wawancara Klinis DAIC-WOZ. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026221177-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Kesehatan mental, khususnya depresi, merupakan isu kesehatan publik yang signifikan dan memengaruhi kualitas hidup individu, sehingga mendorong kebutuhan akan metode deteksi dini berbasis teknologi yang efisien. Penelitian ini fokus pada pembangunan model klasifikasi teks untuk mendeteksi depresi pada tingkat partisipan menggunakan transkrip wawancara klinis DAIC-WOZ, serta membandingkan performa model baseline klasik TF-IDF + Linear SVM dengan beberapa model transformer pralatih, yaitu BERT, RoBERTa, DistilBERT, dan XLM-R. Sistem klasifikasi ini menggunakan skor PHQ-8 sebagai label referensi, sehingga mampu memetakan partisipan ke kategori depresi dan non-depresi secara otomatis. Metodologi penelitian meliputi identifikasi permasalahan dan studi literatur, akuisisi serta pra-pemrosesan transkrip DAIC-WOZ untuk menghasilkan korpus yang bersih dan terstruktur, pembentukan dokumen tingkat partisipan, pembangunan model baseline TF-IDF + Linear SVM serta model transformer pralatih, evaluasi komparatif menggunakan split resmi DAIC-WOZ, dan analisis kinerja dengan metrik seperti akurasi, precision, recall, F1-score, macro-F1, dan balanced accuracy. Selain itu, dilakukan eksplorasi pola linguistik antar kelas menggunakan analisis n-gram pasca-evaluasi untuk memberikan insight tambahan tentang perbedaan penggunaan kata atau frasa antara partisipan depresi dan non-depresi. Hasil penelitian diharapkan memberikan kontribusi teoritis bagi pengembangan NLP di bidang kesehatan mental sekaligus manfaat praktis sebagai landasan pemilihan model yang tepat untuk sistem skrining depresi berbasis teks yang lebih efisien dan mudah diimplementasikan.
======================================================================================================================================
Mental health, particularly depression, is a significant public health concern that affects individuals’ quality of life and motivates the need for more efficient technology-based approaches to early detection. This study focuses on building a text classification model to detect depression at the participant level using clinical interview transcripts from the DAIC-WOZ dataset, and compares the performance of a classical baseline model (TF-IDF + Linear SVM) with several pretrained transformer models, namely BERT, RoBERTa, DistilBERT, and XLM-R. The classification system utilizes PHQ-8 scores as reference labels to automatically categorize participants as depressive or non-depressive. The methodology includes problem identification and literature review, acquisition and preprocessing of DAIC-WOZ transcripts to generate a clean and structured corpus, construction of participant-level documents, development of a TF-IDF + Linear SVM baseline model and pretrained transformer models, comparative evaluation using official DAIC-WOZ splits, and performance analysis with metrics such as accuracy, precision, recall, F1-score, macro-F1, and balanced accuracy. Additionally, post-evaluation n-gram analysis is conducted to explore linguistic patterns between classes, providing insights into differences in word or phrase usage among depressive and non-depressive participants. The study is expected to contribute theoretically to NLP applications in mental health and practically as a basis for selecting appropriate models for more efficient and implemenTabel text-based depression screening systems.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	depresi, pemrosesan bahasa alami, TF-IDF, transformer pralatih, DAIC-WOZ, depression, natural language processing, TF-IDF, pretrained transformers, DAIC-WOZ.
Subjects:	B Philosophy. Psychology. Religion > BF Psychology B Philosophy. Psychology. Religion > BF Psychology > BF318 Learning, Psychology of (Deep learning) Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA76.9.D343 Data mining. Querying (Computer science) R Medicine > R Medicine (General)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Ariq Alwin
Date Deposited:	01 Aug 2026 02:15
Last Modified:	01 Aug 2026 02:15
URI:	http://repository.its.ac.id/id/eprint/141217

Actions (login required)

View Item