Traditional cardiovascular diagnostic procedures are often time- and resource-intensive, motivating the need for more efficient machine learning-based approaches. This work introduces a new Python toolbox for data preprocessing. The new toolbox is named pyPSG and enables the analysis of multimodal physiological signals. It supports ECG, PPG, HRV, and SpO2 signals, and its modular design allows easy extension to other signals. It is important to note that there was already a toolbox for each signal, with its own validated biomarkers. The pyPSG integrates them in one environment, resulting in a collection of 173 standardized, validated biomarkers. Using the toolbox, an ML algorithm was trained on the public dataset of the 2015 CinC Challenge, whose goal was to distinguish between true and false arrhythmia alarms. A total of 750 recordings were processed. Each record is 5 minutes long, the alarm occurring in the last 10 seconds. As a first step, the signals were segmented into 30-second windows. PPG-, ECG-, and HRV-based biomarkers were then extracted from the PPG and ECG signals using the new toolbox. The models were trained and tested in 25 random iterations to asses their performance. Logistic regression, random forest, and XGBoost classifiers were applied. During the analysis of the last window, the XGBoost model achieved the best F1-score of 0.80. Compared to this, the top two algorithms of the 2015 CinC Challenge achieved F1-scores of 0.87 and 0.86. When using 5-minute signals, the XGBoost model achieved an F1-score of 0.76, while the top two algorithms achieved 0.63 and 0.71. In conclusion, the algorithm developed using the new pyPSG toolbox did not outperform existing approaches when analyzing short-term segments, but achieved comparable performance. The resulting algorithm represents a generalizable approach without dataset-specific optimization. Furthermore, when analyzing 5-minute signals, it achieved superior results. pyPSG is available at https://pypsg-github.readthedocs.io/en/latest/.