Resumo:
This work proposes a computationally efficient and reproducible approach for the automatic
detection of beats indicative of AVNRT in electrocardiograms of pediatric patients. The
automatic identification of AVNRT in pediatric patients is a subject practically unexplored
in the literature, but it has many benefits, including faster triage, support for emergency
clinical decisions, and enabling continuous monitoring solutions in wearable devices. For
the proposed technique, beats were segmented based on the detection of R peaks, and
temporal and morphological characteristics were extracted from each beat, including classic
descriptors (amplitudes and RR intervals) and Hjorth parameters (Activity, Mobility, and
Complexity), which synthesize information on amplitude, approximate frequency, and
temporal variation with low computational cost. The evaluation was conducted using the
public database Leipzig Heart Center ECG Database, containing records of 13 pediatric
patients diagnosed with AVNRT. Three supervised machine learning models were trained
and compared: Support Vector Machine (SVM), Random Forest, and LightGBM. The
metrics obtained in the tests indicated performance comparable to or superior to the state
of the art in terms of discrimination and balance between precision and sensitivity, with an
F1-score of 0.983 and an AUC of 0.989 in SVM; LightGBM presented a comparable F1-score
(0.986) with lower prediction latency, suggesting an operational advantage for real-time
applications. Compared to previous studies, the results obtained were similar to or superior
to those reported in analogous tasks in predominantly adult populations, indicating the
potential of the proposed low-complexity methodology as an efficient alternative to deep
models, especially in data-scarce scenarios.
