Fusi Multimodal Gate Attention Dengan Class-Aware CutMix Untuk Prediksi Mortalitas Pasien Congestive Heart Failure Di Intensive Care Unit

Pratama, Mohammad Yoga (2026) Fusi Multimodal Gate Attention Dengan Class-Aware CutMix Untuk Prediksi Mortalitas Pasien Congestive Heart Failure Di Intensive Care Unit. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025241029-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Prediksi mortalitas dalam rumah sakit pada pasien Congestive Heart Failure (CHF) di Intensive Care Unit (ICU) menghadapi dua tantangan yang saling berkaitan. Pertama, model prediksi yang ada umumnya hanya memanfaatkan data terstruktur dari Electronic Health Record (EHR) dan mengabaikan konteks klinis yang terkandung dalam catatan naratif dokter dan perawat. Kedua, data tabular klinis pada populasi ini menunjukkan class overlap yang parah, sehingga classifier standar cenderung mengabaikan kelas minoritas, yakni pasien yang meninggal. Untuk mengatasi kedua tantangan tersebut, tesis ini mengusulkan Multimodal Gate Attention Fusion with CutMix Aware (MGAF-CMA), sebuah kerangka kerja terpadu yang memfusikan representasi teks catatan klinis dari Bio_ClinicalBERT dengan representasi tabular hasil regularisasi CutMix asimetris berbasis kelas melalui mekanisme gate attention per-dimensi. Kerangka ini dievaluasi pada kohort MIMIC-III v1.4 sebanyak 13.956 pasien dan mencapai Sensitivity 0,615, F1-score 0,497, AUROC 0,831, dan AUPRC 0,540. Dibandingkan baseline unimodal teks Bio_ClinicalBERT, MGAF-CMA menaikkan Sensitivity sebesar 147% dan AUROC sebesar 9,2%. Analisis interpretabilitas menunjukkan bahwa prediksi didorong oleh skor severitas ICU dan penanda fungsi ginjal, konsisten dengan faktor risiko yang telah diketahui pada literatur klinis mortalitas CHF, sehingga MGAF-CMA menyediakan kerangka prediksi mortalitas dini pasien CHF di ICU yang interpretabel.
======================================================================================================================================
In-hospital mortality prediction for Congestive Heart Failure (CHF) patients in the Intensive Care Unit (ICU) faces two interconnected challenges. First, existing prediction models predominantly rely on structured Electronic Health Record (EHR) data and overlook the clinical context contained in the narrative notes written by physicians and nurses. Second, tabular clinical data in this population exhibits severe class overlap, causing standard classifiers to overlook the minority class of deceased patients. To address these two challenges, this thesis proposes Multimodal Gate Attention Fusion with CutMix Aware (MGAF-CMA), a unified framework that fuses clinical-note text representations from Bio_ClinicalBERT with tabular representations obtained through class-conditional asymmetric CutMix regularisation, integrated via a per-dimension gate attention mechanism. The framework was evaluated on the MIMIC-III v1.4 cohort of 13,956 patients and achieved a Sensitivity of 0.615, F1-score of 0.497, AUROC of 0.831, and AUPRC of 0.540. Compared with a unimodal Bio_ClinicalBERT text baseline, MGAF-CMA improved Sensitivity by 147% and AUROC by 9.2%. Interpretability analysis showed that the predictions were driven by ICU severity scores and renal function markers, consistent with the risk factors established in the clinical literature on CHF mortality, thereby providing an interpretable framework for early mortality prediction in CHF ICU patients.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Class-Aware CutMix, Congestive heart failure, Explainable AI, MGAF- CMA, MIMIC-III, mortalitas ICU, Class-Aware CutMix, Congestive heart failure, Explainable AI, ICU mortality, MGAF-CMA, MIMIC-III
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) R Medicine > R Medicine (General) > R858 Deep Learning
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Mohammad Yoga Pratama
Date Deposited:	30 Jul 2026 06:32
Last Modified:	30 Jul 2026 06:32
URI:	http://repository.its.ac.id/id/eprint/139795

Actions (login required)

View Item