Klasifikasi Kesehatan Kuda Berdasarkan Data Biometrik Menggunakan One-Dimensional Convolutional Neural Network dan Support Vector Machine

Abstract

Kuda memiliki peran penting dalam kehidupan manusia, baik sebagai hewan kerja, transportasi, olahraga, rekreasi, dan pendukung kebutuhan biomedis melalui produksi plasma antiserum. Namun, kesehatan kuda perlu mendapat perhatian serius karena gangguan kesehatan dapat memengaruhi produktivitas, dan kelangsungan hidupnya. Salah satu gangguan yang sering ditemukan dan berisiko tinggi adalah kolik. Kolik dapat menyebabkan perubahan perilaku dan parameter fisiologis, seperti peningkatan detak jantung, perubahan laju respirasi, perubahan suhu tubuh, dan penurunan saturasi oksigen. Oleh karena itu, deteksi dini diperlukan untuk meminimalkan kematian pada kuda. Keterbatasan pengamatan manual mendorong pemanfaatan biosensor dan wearable sensor untuk memantau kondisi fisiologis kuda secara objektif, kontinu, dan berbasis data. Dalam penelitian ini, data biometrik dari Smart Halter versi 1.5 digunakan sebagai dasar klasifikasi kesehatan kuda secara otomatis. Penelitian ini bertujuan membangun model klasifikasi kesehatan kuda berbasis data biometrik menggunakan One-Dimensional Convolutional Neural Network (1D-CNN) dan Support Vector Machine (SVM), serta mengevaluasi kinerjanya berdasarkan performa klasifikasi dan efisiensi waktu komputasi. Data diperoleh dari kuda di Biofarma Cisarua, Bandung dan Equestrian Park IPB University pada 7–8 September 2023, 14 November 2023, serta 16, 19, dan 21 Mei 2025. Data terdiri atas dua dataset dengan struktur fitur berbeda, yaitu dataset biometrik kuda 1 sebanyak 2340 baris dengan fitur detak jantung, suhu tubuh, dan saturasi oksigen berlabel sehat, dan dataset biometrik kuda 2 sebanyak 2009 baris dengan fitur detak jantung, suhu tubuh, dan laju respirasi berlabel sehat dan sakit. Untuk membentuk empat fitur masukan lengkap, yaitu detak jantung, suhu tubuh, saturasi oksigen, dan laju respirasi, dilakukan praproses berupa imputasi menggunakan Random Forest Regressor, penanganan outlier dengan Interquartile Range (IQR), serta normalisasi menggunakan Min-Max Normalization. Pemodelan dilakukan menggunakan dua pendekatan, yaitu 1D-CNN dan SVM, yang keduanya dioptimasi menggunakan Particle Swarm Optimization (PSO). Model 1D-CNN digunakan karena data biometrik memiliki karakteristik satu dimensi dan berpotensi mengandung pola lokal antar fitur yang dapat dipelajari melalui operasi konvolusi. Sementara itu, SVM digunakan sebagai model pembanding karena memiliki kemampuan generalisasi yang baik pada permasalahan klasifikasi. Pada 1D-CNN, hyperparameter yang dioptimasi meliputi jumlah filter, ukuran kernel, jumlah dense layer, jumlah neuron, dropout rate, learning rate, batch size, dan jumlah epoch. Pada SVM, hyperparameter yang dioptimasi meliputi nilai C, gamma, jenis kernel, dan skema decision values. Evaluasi model dilakukan menggunakan accuracy, precision, recall, F1-Score, confusion matrix, precision-recall curve, ROC curve, serta waktu komputasi. Hasil penelitian menunjukkan bahwa kedua model mampu melakukan klasifikasi kesehatan kuda dengan performa tinggi. Model 1D-CNN dengan PSO memberikan hasil terbaik pada iterasi ke-100 dengan accuracy 98,14%, precision 92,55%, recall 98,68%, dan F1-Score 95,51%, menggunakan konfigurasi 32 filter, kernel size 3, 1 dense layer dengan 64 neuron, dropout rate 0,288, learning rate 0,00541, batch size 32, dan 62 epoch. Berdasarkan confusion matrix, model ini berhasil mengklasifikasikan 589 data kuda sehat dan 149 data kuda sakit dengan benar, dengan 12 false positive dan 2 false negative. Sementara itu, model SVM dengan PSO juga menunjukkan performa kompetitif dengan accuracy 97,87%, precision 92,45%, recall 97,35%, dan F1-Score 94,84%, menggunakan hyperparameter terbaik C = 10, gamma = 10, kernel RBF, dan skema decision value one-vs-rest (OVR). Model SVM berhasil mengklasifikasikan 589 data kuda sehat dan 147 data kuda sakit dengan benar, dengan 12 false positive dan 4 false negative. Secara keseluruhan, 1D-CNN lebih unggul terutama pada nilai recall, sehingga lebih relevan untuk kebutuhan deteksi dini kuda sakit. Perbandingan kedua model menunjukkan adanya trade-off antara performa klasifikasi dan efisiensi komputasi. Model 1D-CNN dengan PSO memberikan performa terbaik secara keseluruhan, terutama karena memiliki recall tertinggi sehingga lebih mampu meminimalkan risiko kuda sakit yang tidak terdeteksi. Namun, model ini membutuhkan waktu komputasi yang lebih besar dan bervariasi, bahkan dapat mencapai beberapa jam hingga belasan jam pada iterasi tertentu. Sebaliknya, model SVM dengan PSO membutuhkan waktu komputasi yang lebih singkat, yaitu kurang dari satu jam hingga sedikit di atas satu jam, dengan performa yang tetap kompetitif. Berdasarkan hasil tersebut, 1D-CNN dengan PSO dapat direkomendasikan apabila prioritas utama sistem adalah akurasi dan sensitivitas deteksi, khususnya untuk mengurangi kemungkinan kuda sakit tidak terdeteksi. Sementara itu, SVM dengan PSO dapat menjadi alternatif yang lebih efisien apabila sistem diterapkan pada lingkungan dengan keterbatasan sumber daya komputasi. Secara keseluruhan, penelitian ini menunjukkan bahwa data biometrik dari Smart Halter dapat dimanfaatkan untuk membangun model klasifikasi kesehatan kuda berbasis machine learning dan deep learning. Hasil penelitian ini diharapkan dapat menjadi dasar pengembangan sistem pemantauan kesehatan kuda secara otomatis, terutama untuk mendukung deteksi dini kondisi sakit atau gangguan kesehatan seperti kolik.

Horses play an important role in human life, including as working animals, means of transportation, animals for sports and recreation, and as biomedical resources through the production of plasma for antiserum production. However, horse health requires serious attention because health disorders may affect productivity and survival. One of the most common and high-risk health disorders in horses is colic. Colic may cause changes in behaviour and physiological parameters, such as increased heart rate, altered respiratory rate, changes in body temperature, and decreased oxygen saturation. Therefore, early detection is needed to minimize mortality in horses. The limitations of manual observation have encouraged the use of biosensors and wearable sensors to monitor horses’ physiological conditions objectively, continuously, and based on data. In this study, biometric data obtained from Smart Halter version 1.5 were used as the basis for automatic horse health classification. This study aimed to develop a horse health classification model based on biometric data using One-Dimensional Convolutional Neural Network (1D-CNN) and Support Vector Machine (SVM), as well as to evaluate their performance based on classification results and computational efficiency. The data were obtained from horses at Biofarma Cisarua, Bandung, and Equestrian Park IPB University on 7–8 September 2023, 14 November 2023, and 16, 19, and 21 May 2025. The data consisted of two datasets with different feature structures. The first horse biometric dataset contained 2,340 rows with heart rate, body temperature, and oxygen saturation features, and was labelled as healthy. The second horse biometric dataset contained 2,009 rows with heart rate, body temperature, and respiratory rate features, and was labelled as healthy and sick. To construct four complete input features, namely heart rate, body temperature, oxygen saturation, and respiratory rate, preprocessing was performed through imputation using Random Forest Regressor, outlier handling using the Interquartile Range method, and normalization using Min-Max normalization. Modelling was conducted using two approaches, namely 1D-CNN and SVM, both of which were optimized using Particle Swarm Optimization (PSO). The 1D-CNN model was used because biometric data have one-dimensional characteristics and may contain local patterns among features that can be learned through convolution operations. Meanwhile, SVM was used as a comparison model because it has good generalization ability in classification problems. In the 1D-CNN model, the optimized hyperparameters included the number of filters, kernel size, number of dense layers, number of neurons, dropout rate, learning rate, batch size, and number of epochs. In the SVM model, the optimized hyperparameters included C, gamma, kernel type, and decision value scheme. Model evaluation was conducted using accuracy, precision, recall, F1-Score, confusion matrix, precision-recall curve, ROC curve, and computational time. The results showed that both models were able to classify horse health status with high performance. The PSO-optimized 1D-CNN model achieved the best result at the 100th iteration, with an accuracy of 98.14%, precision of 92.55%, recall of 98.68%, and F1-Score of 95.51%. This result was obtained using a configuration of 32 filters, a kernel size of 3, one dense layer with 64 neurons, a dropout rate of 0.288, a learning rate of 0.00541, a batch size of 32, and 62 epochs. Based on the confusion matrix, this model correctly classified 589 healthy horse data and 149 sick horse data, with 12 false positives and 2 false negatives. Meanwhile, the PSO-optimized SVM model also showed competitive performance, with an accuracy of 97.87%, precision of 92.45%, recall of 97.35%, and F1-Score of 94.84%. The best hyperparameters for this model were C = 10, gamma = 10, RBF kernel, and one-vs-rest (OVR) decision value scheme. The SVM model correctly classified 589 healthy horse data and 147 sick horse data, with 12 false positives and 4 false negatives. Overall, 1D-CNN performed better, particularly in terms of recall, making it more relevant for the early detection of sick horses. The comparison between the two models indicated a trade-off between classification performance and computational efficiency. The PSO-optimized 1D-CNN model provided the best overall performance, mainly because it achieved the highest recall, which made it more capable of minimizing the risk of sick horses being left undetected. However, this model required greater and more variable computational time, reaching several hours to more than ten hours in certain iterations. In contrast, the PSO-optimized SVM model required shorter computational time, ranging from less than one hour to slightly more than one hour, while still maintaining competitive performance. Based on these results, the PSO-optimized 1D-CNN model can be recommended when the main priority of the system is classification accuracy and detection sensitivity, particularly to reduce the possibility of sick horses being misclassified as healthy. Meanwhile, the PSO-optimized SVM model can serve as a more efficient alternative when the system is implemented in an environment with limited computational resources. Overall, this study demonstrates that biometric data obtained from Smart Halter can be used to develop horse health classification models based on machine learning and deep learning. The findings of this study are expected to serve as a foundation for the development of an automatic horse health monitoring system, particularly to support the early detection of illness or health disorders such as colic.