Background Arrhythmic tendency has been paradoxically associated with intensive exercise in both human and equine athletes. Our previous research has indicated that analysis of electrocardiogram (ECG) complexity might be a valuable tool for early arrhythmia detection in humans. We aimed to determine if ambulatory equine ECG recorded by in vivo telemetric electrocardiography might be amenable to semi-automated complexity analysis. Methods Non-invasive ECG recordings were collected during routine clinical work up from 51 healthy horses using telemetric ECG recorder. No ethical approval was required as no recordings were performed specifically for this study. The study subjects were participating in race training and none showed clinically significant cardiac abnormalities on prior routine examination. Data were recorded at 500 Hz sampling rate in continuous episodes, before and during a period of acceleration from walk to canter. The artefact-free segments of >60 seconds were resampled to 125 Hz rate and filtered using the digital band-pass Butterworth filter (0.13 Hz to 21 Hz). Filtered recordings were converted to binary strings using a custom beat detection algorithm. Only the strips where beat detection was 100% correct were used for subsequent complexity analysis. Binary beat-detection strings were subjected to complexity analysis using several previously published methods: Lempel-Ziv ’76 [LZ76] (1), Lempel-Ziv ’78 [LZ78](2), Titchener T-complexity[TC] (3) and Croll’s BiEntropy [CE] (4). Results To estimate the minimum duration of ECG required for the reliable complexity analysis we evaluated the dependence of complexity of ECG strip on its duration. Three 120-sec lead I ECG strips displaying stable heart rate were used to generate individual sets of random sub-strips of different duration (20, 24, 28, 35, 42, 50, 60, 72, 86 and 100 sec; 60 sub-strips each). It was found that LZ78 and CE values are progressively decreasing with the increase of analysed strip length while LZ76 and TC values are asymptotically closing to the “true value”. Variation coefficients for both LZ76 and TC decreased from 10-12% for 20-sec strips to 4-5% for strips longer than 72 sec. Discussion Our results suggest that Lempel-Ziv 76 or Titchener T-complexity, but not LZ78 or Croll’s BiEntropy might be the feasible tools to estimate the ECG recording complexity. Further study will be needed to improve the performance of complexity analysis and decrease the errors due to random sampling.