Multimodal Fusion dengan Pendekatan Deep Learning pada Sistem Pendeteksi Risiko Komplikasi Continuous Ambulatory Peritoneal Dialysis (CAPD)

Andhina, Jayanti Totti (2024) Multimodal Fusion dengan Pendekatan Deep Learning pada Sistem Pendeteksi Risiko Komplikasi Continuous Ambulatory Peritoneal Dialysis (CAPD). Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025201037-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (2MB) | Request a copy

Abstract

Continuous Ambulatory Peritoneal Dialysis (CAPD) merupakan alternatif terapi dialisis peritoneal yang dapat dimanfaatkan secara optimal karena pasien dapat melakukannya sendiri di rumah dan tidak memerlukan sarana transportasi menuju pusat hemodialisis. Di sisi lain, CAPD juga memiliki kekurangan berupa risiko infeksi yang tinggi akibat kegagalan membran peritoneum yang disebabkan oleh kelalaian pasien dan pengoperasian yang tidak sesuai standar. Kelainan warna pada cairan effluent dialysate dan kondisi klinis pasien dapat menjadi indikator awal untuk mendeteksi adanya risiko komplikasi pada pasien CAPD. Sebelumnya, telah terdapat penelitian yang memanfaatkan deep learning image classification untuk mendeteksi risiko komplikasi CAPD menggunakan citra effluent dialysate. Pada penelitian ini, ditambahkan faktor kondisi klinis pasien yang didapatkan dari aplikasi SahabatCAPD dan berbentuk data tabular. Kedua indikator (gambar dan tabular) kemudian digabungkan sehingga menghasilkan model deep learning dan machine learning baru yang dapat mendeteksi risiko komplikasi pada pasien CAPD. Metode deep learning dengan transfer learning fine tuning digunakan untuk menghasilkan model baru untuk task klasifikasi citra effluent dialysate dari beberapa model pre-trained: YOLOv8, MobileNetV3, ConvNeXTV2, dan MobileViTv2. Selanjutnya, model deep learning terbaik digunakan untuk ekstraksi fitur dari setiap gambar. Hasil ekstraksi fitur berupa single feature vector dengan dimensi 512 dan 2 yang didapatkan dari fully connected layer akan digabungkan dengan data klinis pasien, seperti jenis cairan, volume masuk, volume ultrafiltrasi, dan dwell time. Penggabungan data pada penelitian ini menggunakan pendekatan multimodal early fusion. Data-data yang telah diseleksi dalam bentuk tabular selanjutnya akan dijadikan input pada model machine learning: logistic regression, k-nearest neighbors (KNN), support vector machine, decision tree, random forest, dan AdaBoost. Hasil penelitian menunjukkan bahwa model deep learning terbaik dicapai oleh MobileNetV3 dengan recall 86% dan F1-score 92% dengan praproses augmentasi. Sedangkan untuk model machine learning terbaik dicapai oleh random forest dan AdaBoost 95% dan F1-score sebesar 98% pada skenario penggunaan 2 fitur ekstraksi gambar dan 4 fitur data tabular tanpa oversampling. Dari hasil tersebut, kedua model diharapkan dapat membantu dalam hal deteksi risiko komplikasi CAPD pada SahabatCAPD sehingga pasien dan tenaga medis dapat menentukan tindakan lebih dini apabila ditemukan risiko komplikasi yang serius.
====================================================================================================================================
Continuous Ambulatory Peritoneal Dialysis (CAPD) is an alternative peritoneal dialysis therapy that can be optimally utilized because patients can perform it themselves at home, eliminating the need for transportation to a hemodialysis center. However, CAPD also has the drawback of a high risk of infection due to peritoneal membrane failure caused by patient negligence and improper operation. Discoloration of the effluent dialysate and the patient's clinical condition can be early indicators for detecting the risk of complications in CAPD patients. Previously, there have been studies utilizing deep learning image classification to detect CAPD complication risks using effluent dialysate images. In this study, the patient's clinical condition factors obtained from the SahabatCAPD application in tabular data form are added. Both indicators (images and tabular data) are then combined to produce a new deep learning and machine learning model that can detect complication risks in CAPD patients. The deep learning method with transfer learning fine-tuning is used to create a new model for the effluent dialysate image classification task from several pre-trained models: YOLOv8, MobileNetV3, ConvNeXTV2, and MobileViTv2. The best deep learning model is then used to extract features from each image. The feature extraction results in a single feature vector with dimensions 512 and 2 obtained from the fully connected layer will be combined with the patient's clinical data, such as fluid type, input volume, ultrafiltration volume, and dwell time. Data combination in this study uses the multimodal early fusion approach. The selected tabular data will then be used as input to machine learning models: logistic regression, k-nearest neighbors (KNN), support vector machine, decision tree, random forest, and AdaBoost. The study results show that the best deep learning model is achieved by MobileNetV3 with a recall of 86% and an F1-score of 92% with augmentation preprocessing. Meanwhile, the best machine learning models are achieved by random forest and AdaBoost with 95% recall and an F1-score of 98% in the scenario using 2 image extraction features and 4 tabular data features without oversampling. From these results, both models are expected to assist in detecting CAPD complication risks in SahabatCAPD so that patients and medical personnel can take early action if serious complication risks are found.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Continuous Ambulatory Peritoneal Dialysis (CAPD), deep learning, effluent dialysate, machine learning, multimodal fusion, Continuous Ambulatory Peritoneal Dialysis (CAPD), deep learning, effluent dialysate, machine learning, multimodal fusion
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T174 Technological forecasting T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Jayanti Totti Andhina
Date Deposited:	01 Aug 2024 03:04
Last Modified:	12 Sep 2024 06:02
URI:	http://repository.its.ac.id/id/eprint/110204

Actions (login required)

View Item