Aims: Photoplethysmography (PPG) is increasingly used for wearable cardiac monitoring. With increasing interest in early detection of cardiac arrest (CA) with wearables, PPG data containing CA events are essential for developing and validating CA detection algorithms. However, these data are scarce. In this work, we aimed to compile and publicly share a PPG dataset of CA events. Methods: We obtained data from the MIMIC-III Clinical Database and Waveform Database Matched Subset, containing patient monitor data from 10,282 patients admitted to the ICU. Records were initially selected based on the patient's age (at least 18 years old), and the presence of (1) a PPG signal, (2) either ECG lead II or continuous arterial blood pressure (ABP), and (3) automated heart rhythm annotations by the patient monitors. We identified approximate timestamps of potential cardiac arrests by searching for heart rhythms classified by the patient monitors as asystole or ventricular tachycardia/fibrillation, which are frequently occurring types of CA. False monitor alarms were discarded by manual review of the ECG and ABP around the CA event timestamp. As patient monitor heart rhythm annotations are not available for many records in the MIMIC-III Waveform Database, future work includes developing alternative automated annotation tools to assist manual review. Furthermore, supplementation of the dataset with verified negative samples (not containing cardiac arrest events) will allow for testing the specificity of detection algorithms. Results: After initial selection, 12,014 records from 6,461 distinct patients were found. In these records, 233 potential CA events were detected by the patient monitor. Manual review confirmed 24 of these potential events as true cardiac arrest events captured by PPG. Conclusion: The use of patient monitor alarm-guided searching yields 24 PPG recordings containing CA events from the MIMIC-III Databases, which could be used for the development and validation of CA detection algorithms.