Análise da Classificação de Habilidade em Punção Lombar: Uma Abordagem Baseada em Sensores Inerciais e Aprendizado de Máquina

Abstract

Lumbar puncture (LP) is an invasive procedure performed to collect cerebrospinal fluid and is commonly used in neurological practice for diagnostic indications and, in some cases, therapeutic purposes. Effective collection depends on the practitioner maintaining a stable hand throughout the procedure; otherwise, complications such as pain, nerve injury, infections, and hemorrhages may occur. The objective of this study is to analyze signals captured by an inertial sensor during the performance of a lumbar puncture procedure on a low-fidelity mannequin and to compare them between participant groups. Method: This study included 26 volunteers, divided into four groups (G1, G2, G3, and G4) composed of medical graduates and senior medical students. Data collection was performed using an inertial sensor (comprising an accelerometer, a gyroscope, and a magnetometer) attached to the participants’ hands, who were wearing surgical gloves. The movement involved simulating the lumbar puncture procedure on a low-fidelity mannequin. Eighteen features were extracted from the collected signals, related to signal amplitude, frequency, entropy, variability, and distribution shape. Group comparisons were conducted using Post-Hoc Kruskal-Wallis tests, and sensitivity, specificity, precision, and accuracy were calculated for the classifiers used: Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Naive Bayes (NB). Results: The Post-Hoc Kruskal-Wallis test showed that G3, composed of students without prior LP experience, had the lowest performance, as expected, while G2, composed of physicians who perform LP less frequently, also showed lower performance. Regarding the classifiers, the highest sensitivity was observed in the KNN classifier (95.7%) for the G1/G2 comparison and in the KNN classifier (88.1%) for the G2/G3 comparison; specificity was highest in the KNN classifier (98.4%) for the G2/G4 comparison and in the SVM classifier (78.8%) for the G2/G3 comparison; precision was highest in the KNN classifier (98.8%) and (89%) in the G2/G4 comparison; and accuracy (89%) was highest in the KNN classifier for the G1/G2 comparison.. Conclusion: The use of inertial sensors proved to be a cost-effective alternative to the analysis proposed in this study. The techniques applied for data analysis demonstrated effectiveness in evaluating the collected signals.