Federated Learning–Bidirectional Long Short-Term Memory-Based Model Untuk Peramalan Jumlah Kasus Dengue

Fitriyan, Muhammad Reyhan (2024) Federated Learning–Bidirectional Long Short-Term Memory-Based Model Untuk Peramalan Jumlah Kasus Dengue. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026201148-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 October 2026. Download (3MB) | Request a copy

Abstract

Dengue merupakan penyakit menular yang mengancam nyawa penduduk global terutama yang tinggal di wilayah endemis dengue, termasuk Indonesia. Jumlah kasus dengue di Indonesia masih melampaui target. Salah satu upaya yang dapat dilakukan untuk mengurangi penyebaran dan dampak infeksi dengue adalah pengembangan model peramalan jumlah kasus. Federated learning (FL) sudah banyak diimplementasikan pada bidang medis karena dapat menjaga privasi data dan meningkatkan performa model pada beberapa kasus. Namun, belum ditemukan model peramalan jumlah kasus dengue yang dikembangkan dengan pendekatan FL. Oleh karena itu, model peramalan jumlah kasus dengue berbasis bidirectional long short-term memory (BiLSTM) dengan pendekatan FL diajukan. Data seluruh variabel masukan untuk masing-masing puskesmas di Kabupaten Malang yang telah dikumpulkan diterapkan berbagai metode praproses sesuai kebutuhan dan karakteristik data. Kemudian, pelatihan model BiLSTM secara individual dengan penyetelan hiperparameter dilakukan untuk dua model dasar dengan jumlah lapis tersembunyi yang berbeda, yaitu dua dan empat. Kedua konfigurasi model terbaik berdasarkan rata-rata terbobot RMSE digunakan sebagai model lokal pada proses pelatihan dengan pendekatan FL dan juga sebagai pembanding performa model FL. FL dilakukan dengan menugaskan subset klien puskesmas untuk melatih model pada data lokal masing-masing, yang kemudian hasilnya dikirimkan ke server untuk diagregasi untuk membentuk model global. Berbagai skenario yang memvariasikan jumlah epoch lokal, ronde komunikasi, dan klien partisipan diterapkan untuk masing-masing model. Untuk model BiLSTM dengan dua lapis tersembunyi, skenario FL terbaiknya menghasilkan rata-rata terbobot RMSE sebesar 2,588, yang lebih rendah 2,96% dari hasil pendekatan individual. Sementara itu, untuk model BiLSTM dengan empat lapis, dihasilkan rata-rata terbobot RMSE sebesar 2,582, yang lebih rendah 5,80% daripada hasil pendekatan individual. Peningkatan juga terlihat dalam kemampuan model untuk menghasilkan ramalan yang mengikuti pola dan fluktuasi data aktual klien puskesmas. Kedua model FL tersebut secara keseluruhan juga memiliki performa lebih baik dibandingkan model FL dengan pengklasteran klien berdasarkan ketinggian dataran wilayah puskesmas.
========================================================================================================================
Dengue is a life-threatening infectious disease that poses a threat to the global population, especially those living in dengue-endemic areas, including Indonesia. The number of dengue cases in Indonesia still exceeds the target. One effort to reduce the spread and impact of dengue infection is the development of a case forecasting model. Federated learning (FL) has been widely implemented in the medical field because it can protect data privacy and improve model performance in several cases. However, a dengue case number forecasting model has not yet been developed using FL approach. Therefore, a dengue case number forecasting model based on bidirectional long short-term memory (BiLSTM) with FL approach is proposed. Data for all input variables for each puskesmas (public health center) in Malang Regency are preprocessed using various methods according to the characteristics of the data. Then, two BiLSTM models with two and four hidden layers were individually trained with hyperparameter tuning. The best configurations of each model based on the weighted average RMSE were used as local models in the FL training process. FL was conducted by assigning subsets of puskesmas clients to train models on their local data, which were then sent to the server for aggregation and formation of the global model. Various scenarios varying the number of local epochs, communication rounds, and participating clients were applied for each model. For the BiLSTM model with two hidden layers, the best FL scenario produced a weighted average RMSE of 2.588, which is 2.96% lower than the individual approach. Meanwhile, for the BiLSTM model with four hidden layers, a weighted average RMSE of 2.582 was achieved, which is 5.80% lower than the individual approach. Improvement was also seen in the models’ ability to produce forecasts that follow the patterns of the actual data of the puskesmas clients. Both FL models overall also performed better than the FL models formed with client clustering based on the altitude of the puskesmas area.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Federated learning, dengue, peramalan, bidirectional long short-term memory, federated learning, dengue, forecasting, bidirectional long short-term memory
Subjects:	H Social Sciences > HA Statistics > HA30.3 Time-series analysis Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. 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 T Technology > T Technology (General)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Reyhan Fitriyan
Date Deposited:	02 Aug 2024 02:43
Last Modified:	02 Aug 2024 02:43
URI:	http://repository.its.ac.id/id/eprint/112129

Actions (login required)

View Item