Limited access to blood tests in underrepresented regions, such as parts of South America, highlights the need for cost-effective and non-invasive methods to identify Chagas disease (CD). CD, caused by the protozoan \textit{Trypanosoma cruzi}, is a neglected tropical disease affecting an estimated six million people worldwide. Efficient use of limited diagnostic resources requires approaches that reduce false positives while increasing the detection of true positive cases. To address this need, the PhysioNet/CinC Challenge 2025 was organized to detect CD using 12-lead ECG signals, leveraging the fact that CD can cause cardiac abnormalities detectable in ECG waveforms. In this study, as part of the PhysioNet 2025 Challenge, we developed a CNN-based lead-wise feature learning model for CD detection. Our team, PhysioWinn, achieved a challenge score of 0.326 on the official leaderboard, ranking 22nd out of 67 teams. In addition, we performed a comprehensive statistical analysis to assess feature- and lead-level importance, revealing that RR Interval RMSSD was significant across all leads and that the most discriminative features were concentrated in the precordial (anterior chest) leads. These findings suggest that targeted feature engineering in precordial leads may further improve CD detection in future work.