"Penelitian ini mengembangkan model estimasi kadar hemoglobin non-invasif menggunakan sinyal photoplethysmography (PPG) tiga panjang gelombang (hijau, inframerah, merah) dari sensor MAXM86161. Data dikumpulkan dari 52 subjek (pria, wanita nonhamil, wanita hamil), dengan 49 fitur diekstraksi dari setiap sinyal. Fitur diseleksi menggunakan enam metode (Correlation, Corr+ReliefF, Lasso, MI, MI+Lasso, ReliefF+MI) dan diuji pada empat model regresi (Random Forest, XGBoost, SVR, MLP). Kombinasi ReliefF+MI dengan 15 fitur menghasilkan performa terbaik pada Random Forest (MAE test 0,49 g/dL, R² test 0,93) dan XGBoost (MAE test 0,51 g/dL, R² test 0,93). Fitur utama meliputi ch1_h1_power dan ch3_dyn_component dari kanal hijau dan merah, serta variabel demografis (usia, gender, trimester), sedangkan kanal inframerah kurang berkontribusi. SVR menghasilkan MAE test >1,84 g/dL dan R² test <0,20, sementara MLP menunjukkan overfitting (R² val < -1,18). Prediksi mencapai akurasi 97% namun kurang akurat pada hemoglobin ekstrem (<12 g/dL atau >16 g/dL, Absolute Error ≥1,2 g/dL). Rekomendasi meliputi perluasan dataset dan optimasi regularisasi model untuk meningkatkan generalisasi sebelum penerapan klinis.
This study developed a model for estimating non-invasive hemoglobin levels using a three-wavelength photoplethysmography (PPG) signal (green, infrared, red) from a MAXM86161 sensor. Data were collected from 52 subjects (men, non-pregnant women, pregnant women), with 49 features extracted from each signal. Features were selected using six methods (Correlation, Corr+ReliefF, Lasso, MI, MI+Lasso, ReliefF+MI) and tested on four regression models (Random Forest, XGBoost, SVR, MLP). The combination of ReliefF+MI with 15 features yields the best performance on Random Forest (MAE test 0.49 g/dL, R² test 0.93) and XGBoost (MAE test 0.51 g/dL, R² test 0.93). Key features include the ch1_h1_power and ch3_dyn_component of the green and red channels, as well as demographic variables (age, gender, trimester), while the infrared channel contributes less. The SVR produced an MAE test >1.84 g/dL and an R² test <0.20, while the MLP showed overfitting (R² val < -1.18). The prediction achieved 97% accuracy but was less accurate at extreme hemoglobin (<12 g/dL or >16 g/dL, Absolute Error ≥1.2 g/dL). Recommendations include expanding the dataset and optimizing the regularization of the model to improve generalization before clinical application. "