Klasifikasi kondisi pasien berdasarkan tanda vital menggunakan Bidirectional Long Short-Term Memory

INDONESIA:

Klasifikasi perburukan kondisi pasien di ICU maupun UGD masih menjadi tantangan karena sistem monitoring yang ada saat ini cenderung bersifat reaktif dan berfokus pada peringatan ambang batas normal. Penelitian ini mengusulkan model klasifikasi kondisi pasien yang proaktif berbasis Bidirectional Long Short-Term Memory dengan menggunakan data time-series tanda vital multivariat, yang meliputi detak jantung, laju pernapasan, tekanan darah, saturasi oksigen, dan suhu tubuh, serta tambahan atribut turunannya. Data bersumber dari Kaggle dan diproses secara sistematis melalui Min-Max Scaling, perhitungan atribut temporal tambahan, dan penanganan batas sekuens pasien. Pengujian dilakukan pada tiga skenario pembagian data yang berbeda. Hasil evaluasi menunjukkan bahwa Skenario A (rasio 80:10:10) selama 50 epoch memberikan performa terbaik dan paling tangguh, dengan tingkat Akurasi mencapai 88,57%, Presisi 90,32%, Recall 86,42%, dan F1-Score 88,33%. Kesimpulannya, arsitektur BiLSTM terbukti andal dalam mendeteksi pola temporal perburukan pasien secara dini dan akurat dengan meminimalkan kesalahan deteksi, sehingga sangat layak untuk diimplementasikan guna mendukung pengambilan keputusan klinis pada sistem pemantauan vital secara real-time.

ENGLISH:

Clasification of patient deterioration in the ICU or emergency department remains a challenge due to current monitoring systems that tend to be reactive and focus solely on normal threshold alerts. This study proposes a proactive patient condition classification model based on Bidirectional Long Short-Term Memory (BiLSTM) using multivariate time-series vital sign data, encompassing heart rate, respiratory rate, blood pressure, oxygen saturation, and body temperature, along with their derived attributes. Data was sourced from Kaggle and systematically processed through Min-Max Scaling, calculation of additional temporal features, and sequence boundary handling. The testing was conducted across three different data-splitting scenarios. Evaluation results demonstrated that Scenario A (80:10:10 ratio) trained over 50 epochs yielded the best and most robust performance, achieving an Accuracy of 88.57%, Precision of 90.32%, Recall of 86.42%, and F1-Score of 88.33%. In conclusion, the BiLSTM architecture proved reliable in accurately detecting early temporal patterns of patient deterioration while minimizing false detection rates (false positives), making it highly feasible for implementation to support clinical decision-making in real-time vital monitoring systems.

ARABIC:

ﻻ ﻳﺰﺍﻝ ﺗﺼﻨﻴﻒ ﺗﺪﻫﻮﺭ ﺣﺎﻟﺔ ﺍﻟﻤﺮﺿﻰ ﻓﻲ ﻭﺣﺪﺓ ﺍﻟﻌﻨﺎﻳﺔ ﺍﻟﻤﺮﻛﺰﺓ (ICU) ﺃﻭ ﻗﺴﻢ ﺍﻟﻄﻮﺍﺭﺉ (UGD) ﻳﻤﺜﻞ ﺗﺤﺪﻳﺎً ﻷﻥ ﺃﻧﻈﻤﺔ ﺍﻟﻤﺮﺍﻗﺒﺔ ﺍﻟﺤﺎﻟﻴﺔ ﺗﻤﻴﻞ ﺇﻟﻰ ﺃﻥ ﺗﻜﻮﻥ ﺗﻔﺎﻋﻠﻴﺔ ﻭﺗﺮﻛﺰ ﻋﻠﻰ ﺍﻹﻧﺬﺍﺭﺍﺕ ﻋﻨﺪ ﺗﺠﺎﻭﺯ ﺍﻟﺤﺪﻭﺩ ﺍﻟﻄﺒﻴﻌﻴﺔ. ﺗﻘﺘﺮﺡ ﻫﺬﻩ ﺍﻟﺪﺭﺍﺳﺔ ﻧﻤﻮﺫﺟﺎً ﺍﺳﺘﺒﺎﻗﻴﺎً ﻟﺘﺼﻨﻴﻒ ﺣﺎﻟﺔ ﺍﻟﻤﺮﻳﺾ ﻳﻌﺘﻤﺪ ﻋﻠﻰ ﺷﺒﻜﺔ ﺍﻟﺬﺍﻛﺮﺓ ﺍﻟﻄﻮﻳﻠﺔ ﻭﺍﻟﻘﺼﻴﺮﺓ ﺍﻟﻤﺪﻯ ﺛﻨﺎﺋﻴﺔ ﺍﻻﺗﺠﺎﻩ Memory) Short-Term Long (Bidirectional ﺑﺎﺳﺘﺨﺪﺍﻡ ﺑﻴﺎﻧﺎﺕ ﺍﻟﺴﻼﺳﻞ ﺍﻟﺰﻣﻨﻴﺔ ﺍﻟﻤﺘﻌﺪﺩﺓ ﺍﻟﻤﺘﻐﻴﺮﺍﺕ ﻟﻌﻼﻣﺎﺕ ﺍﻟﺤﻴﺎﺓ ﺍﻟﺤﻴﻮﻳﺔ، ﻭﺍﻟﺘﻲ ﺗﺸﻤﻞ ﻣﻌﺪﻝ ﺿﺮﺑﺎﺕ ﺍﻟﻘﻠﺐ، ﻭﻣﻌﺪﻝ ﺍﻟﺘﻨﻔﺲ، ﻭﺿﻐﻂ ﺍﻟﺪﻡ، ﻭﺗﺸﺒﻊ ﺍﻷﻛﺴﺠﻴﻦ، ﻭﺩﺭﺟﺔ ﺣﺮﺍﺭﺓ ﺍﻟﺠﺴﻢ، ﺑﺎﻹﺿﺎﻓﺔ ﺇﻟﻰ ﺍﻟﺴﻤﺎﺕ ﺍﻟﻤﺸﺘﻘﺔ ﻣﻨﻬﺎ. ﺗﻢ ﺍﻟﺤﺼﻮﻝ ﻋﻠﻰ ﺍﻟﺒﻴﺎﻧﺎﺕ ﻣﻦ ﻣﻨﺼﺔ Kaggle ﻭﻣﻌﺎﻟﺠﺘﻬﺎ ﺑﺸﻜﻞ ﻣﻨﻬﺠﻲ ﻣﻦ ﺧﻼﻝ ﺗﺤﻮﻳﻞ ﺍﻟﻨﻄﺎﻕ Scaling) (Min-Max، ﻭﺣﺴﺎﺏ ﺍﻟﺴﻤﺎﺕ ﺍﻟﺰﻣﻨﻴﺔ ﺍﻹﺿﺎﻓﻴﺔ، ﻭﻣﻌﺎﻟﺠﺔ ﺣﺪﻭﺩ ﺗﺴﻠﺴﻞ ﺍﻟﻤﺮﺿﻰ. ﻭﺃﺟُﺮﻳﺖ ﺍﻻﺧﺘﺒﺎﺭﺍﺕ ﻋﻠﻰ ﺛﻼﺛﺔ ﺳﻴﻨﺎﺭﻳﻮﻫﺎﺕ ﻣﺨﺘﻠﻔﺔ ﻟﺘﻘﺴﻴﻢ ﺍﻟﺒﻴﺎﻧﺎﺕ. ﺃﻇﻬﺮﺕ ﻧﺘﺎﺋﺞ ﺍﻟﺘﻘﻴﻴﻢ ﺃﻥ ﺍﻟﺴﻴﻨﺎﺭﻳﻮ »ﺃ« )ﺑﻨﺴﺒﺔ (80:10:10 ﻋﻠﻰ ﻣﺪﺍﺭ 50 ﺩﻭﺭﺓ (epoch) ﻗﺪﻡ ﺍﻷﺩﺍﺀ ﺍﻷﻓﻀﻞ ﻭﺍﻷﻛﺜﺮ ﻣﻮﺛﻮﻗﻴﺔ، ﺣﻴﺚ ﺑﻠﻐﺖ ﻧﺴﺒﺔ ﺍﻟﺪﻗﺔ %88,57، ﻭﺍﻟﺪﻗﺔ ﺍﻟﻨﻮﻋﻴﺔ %90,32، ﻭﻣﻌﺪﻝ ﺍﻻﺳﺘﺮﺟﺎﻉ %86,42، ﻭﻣﺆﺷﺮ 88,33 .%F1-Score ﻭﺧﺘﺎﻣﺎً، ﺃﺛﺒﺘﺖ ﺑﻨﻴﺔ BiLSTM ﻣﻮﺛﻮﻗﻴﺘﻬﺎ ﻓﻲ ﺍﻟﻜﺸﻒ ﺍﻟﻤﺒﻜﺮ ﻭﺍﻟﺪﻗﻴﻖ ﻋﻦ ﺍﻷﻧﻤﺎﻁ ﺍﻟﺰﻣﻨﻴﺔ ﻟﺘﺪﻫﻮﺭ ﺣﺎﻟﺔ ﺍﻟﻤﺮﺿﻰ ﻣﻊ ﺗﻘﻠﻴﻞ ﺃﺧﻄﺎﺀ ﺍﻟﻜﺸﻒ ﺇﻟﻰ ﺍﻟﺤﺪ ﺍﻷﺩﻧﻰ، ﻣﻤﺎ ﻳﺠﻌﻠﻬﺎ ﺟﺪﻳﺮﺓ ﺑﺎﻟﺘﻄﺒﻴﻖ ﻟﺪﻋﻢ ﺍﺗﺨﺎﺫ ﺍﻟﻘﺮﺍﺭﺍﺕ ﺍﻟﺴﺮﻳﺮﻳﺔ ﻓﻲ ﺃﻧﻈﻤﺔ ﻣﺮﺍﻗﺒﺔ ﺍﻟﻌﻼﻣﺎﺕ ﺍﻟﺤﻴﻮﻳﺔ ﻓﻲ ﺍﻟﻮﻗﺖ ﺍﻟﻔﻌﻠﻲ.