Sistem Pendukung Keputusan Klinis (CDSS) pada Gangguan Trombosit Berbasis Complete Blood Count (CBC) Menggunakan Machine Learning dan Explainable AI

Muhammad Wahyu Tri Saputra, Ir. Syukron Abu Ishaq Alfarozi, S.T., Ph.D.

2026 | Skripsi | S1 TEKNIK BIOMEDIS

Gangguan trombosit seperti Immune Thrombocytopenia (ITP), Demam Berdarah Dengue (DBD), dan Trombositosis Esensial memiliki profil hematologi yang dapat tumpang tindih, terutama antara ITP dan DBD pada fase trombositopenia akut. Hal ini menyulitkan diagnosis banding berbasis pemeriksaan Complete Blood Count (CBC) standar di fasilitas kesehatan tingkat pertama yang kekurangan tenaga spesialis. Penelitian ini bertujuan mengembangkan Clinical Decision Support System (CDSS) berbasis arsitektur Tri-Brid untuk mengklasifikasikan Normal, ITP, DBD, dan Trombositosis menggunakan parameter CBC standar. Sistem mengintegrasikan probabilitas machine learning (Pilar I), skor gejala klinis (Pilar II), aturan WHO 2009 dan IWG 2009 (Pilar III), serta Explainable Artificial Intelligence (XAI) berbasis SHAP.

Dataset terdiri atas 393 data primer CBC dari rekam medis retrospektif RSA UGM. Studi ablasi membandingkan 36 kombinasi pipeline yang mencakup empat metode seleksi fitur, tiga metode penskalaan, dan tiga algoritma klasifikasi dengan optimasi GridSearchCV pada empat skenario prapemrosesan, yaitu baseline tanpa imputasi dan penyeimbangan kelas, KNNImputer, KNNImputer dengan SMOTE, dan KNNImputer dengan SMOTE-ENN. Pada test set yang identik, SMOTE hanya menaikkan rata-rata MCC sebesar 0,0071 (tidak signifikan), sedangkan ENN menurunkannya sebesar 0,0351. Dibandingkan listwise deletion, KNNImputer berasosiasi dengan kenaikan rata-rata MCC sebesar 0,0755, tetapi selisih ini mencampur efek imputasi dengan perbedaan data uji. Pipeline yang direkomendasikan, yaitu StandardScaler, KNNImputer, Mutual Information, SMOTE, dan Logistic Regression, memperoleh MCC sebesar 0,9584 (IK 95% 0,899– 1,000) pada hold-out test set dan rata-rata MCC validasi silang sebesar 0,8605. Analisis SHAP pada model final menunjukkan PLT sebagai fitur pembeda utama (total mean absolute SHAP 11,62; 41,4?ri total kontribusi), sedangkan PLR dan MCH berkontribusi terutama pada prediksi ITP, serta MCV dan ABS_EOS pada prediksi DBD. Validasi oleh satu dokter spesialis patologi klinik pada model purwarupa, tanpa informasi gejala, menghasilkan Initial Concordance sebesar 70?n skor CLIX-M sebesar 3,00/4,00 (75,0%). Sistem kemudian diimplementasikan sebagai purwarupa CDSS berbasis web dengan visualisasi SHAP.

Hasil penelitian menunjukkan bahwa integrasi CBC standar, machine learning, dan XAI dalam arsitektur Tri-Brid dapat menjadi pendekatan pendukung untuk diferensiasi gangguan trombosit pada fasilitas kesehatan tingkat pertama.

Platelet disorders such as Immune Thrombocytopenia (ITP), Dengue Hemorrhagic Fever (DHF), and Essential Thrombocythemia may exhibit overlapping hematological profiles, particularly ITP and DHF during the acute thrombocytopenic phase. This condition complicates diagnostic differentiation based on standard Complete Blood Count (CBC) examinations in primary healthcare facilities with limited access to specialists. This study aimed to develop a Clinical Decision Support System (CDSS) based on a Tri-Brid architecture to classify Normal, ITP, DHF, and Thrombocytosis cases using standard CBC parameters. The system integrates machine learning probabilities (Pillar I), clinical symptom scores (Pillar II), WHO 2009 and IWG 2009 rules (Pillar III), and SHAP-based Explainable Artificial Intelligence (XAI). 

The dataset consisted of 393 primary CBC data records obtained from retrospective medical records at RSA UGM. The ablation study compared 36 pipeline combinations comprising four feature selection methods, three scaling methods, and three classification algorithms optimized using GridSearchCV under four preprocessing scenarios: a baseline without imputation or class balancing, KNNImputer, KNNImputer with SMOTE, and KNNImputer with SMOTE-ENN. On an identical test set, SMOTE added only 0.0071 to the mean MCC (not significant), whereas ENN reduced it by 0.0351. KNNImputer, compared with listwise deletion, was associated with a 0.0755 increase in mean MCC, but this difference combines the imputation effect with a difference in test data. The recommended pipeline, consisting of StandardScaler, KNNImputer, Mutual Information, SMOTE, and Logistic Regression, achieved an MCC of 0.9584 (95% CI 0.899–1.000) on the hold-out test set and a mean cross-validation MCC of 0.8605. SHAP analysis of the final model identified PLT as the primary discriminative feature (total mean absolute SHAP 11.62; 41.4% of total contribution), while PLR and MCH contributed mainly to ITP prediction, and MCV and ABS_EOS to DHF prediction. Validation by a clinical pathologist on the prototype model, without symptom information, yielded an Initial Concordance of 70% and a CLIX-M score of 3.00/4.00 (75.0%). The system was subsequently implemented as a web-based CDSS prototype featuring SHAP visualization. 

The results indicate that the integration of standard CBC parameters, machine learning, and XAI within a Tri-Brid architecture can serve as a decision-support approach for differentiating platelet disorders in primary healthcare facilities

Kata Kunci : CDSS, gangguan trombosit, Complete Blood Count, machine learning, SHAP