Introduction: Detection of intermittent arrhythmic episodes via long-term cardiac monitoring is important. For instance, atrial fibrillation may be life-threatening in the event of carotid artery embolization and stroke. Several machine-learning models exist to this end, but most are based on single ECG lead-input which may omit important diagnostic information and encumber translation into a hospital setting, where 12-lead ECG is standard. Dimensionality reduction may be achieved via vectorcardiographic (VCG) analysis, whereby the 12-channel ECG signal is converted into three orthonormal leads. Most tools also rely on extracting temporal waveform features, potentially precluding accountancy for long-term physiological variations in heart rate. Generating a 2D ‘attractor’, in which all the data pertaining to the signal is represented in a single bounded space, may cope with these changes. Here, aberrant waveform events are reflected in attractor shape (illustrated in Figure 1) based on their relative, rather than absolute timing within the cardiac cycle and baseline variation is removed without waveform distortion. Aims: Using traces from two online databases, I aimed to combine VCG and attractor analysis to develop a novel tool for real-time, adaptive detection of a range of arrhythmic beat-types, test its performance and then to investigate methods of improving accuracy. Methods: The following is an outline of the proposed programme: Matrix transformation followed by application of Pythagoras’ theorem are applied to convert 12-lead ECG to a single VCG maximum amplitude signal. Pan-Tompkins’ algorithm extracts cycle periods from which 2D attractor co-ordinates are derived. Spectrograms are computed from these values and then used to extract time-frequency moments (instantaneous frequency and spectral entropy). These ‘features’ are employed to train a bidirectional long-short-term memory neural network which classifies single cycles as being arrhythmic or normal. Results: Low training and testing times support the model’s suitability for beat-by-beat analysis and ‘on-the-fly’ re-configuration of network parameters for personalised arrhythmia detection. When trained and validated on patient-specific data, an average classification accuracy of 82.9% (n = 13; SD:14) was achieved. When trained on signals from multiple patients, the model yielded a relatively low accuracy of 63.73% (n = 23; SD: 20.7). However, accuracy was still high for conditions which were better represented in training, including atrial fibrillation (n = 2; mean: 71.45%; SD: 5.59) and paroxysmal supraventricular tachycardia (n = 3; mean: 76.76%; SD: 6.52). When additional data pertaining to Wolff-Parkinson-White syndrome (WPW), an arrhythmia type that had relatively poor representation in training, were added to the training dataset, there was a significant increase in mean cycle classification accuracy in all WPW patients from 53.20% to 77.29% (p < 0.05). Conclusions: This novel amalgamation of VCG and attractor analysis therefore has the capacity to learn meaningful information about ECG signals and has potential utility in automated arrhythmic detection in clinical and day-to-day contexts. The high performance of the bespoke tool suggests its applicability to continuous monitoring of patients with an established risk for further arrhythmic episodes. With more training data, the generically-trained model may be employed for diagnosing arrhythmia in individuals to which the network is naïve.