{
  "abstract": "Introduction Selecting the most appropriate ICD and optimizing device programming to suit individual patient needs remains a significant challenge for healthcare professionals. One common complication associated with newly developed EVICDs is P-wave oversensing, which can lead to inappropriate therapies. In this study, we utilised a machine-learning-based AI tool to analyse P:R ratios over a 24-hour period, aiming to predict P-wave oversensing and provide insights for personalised device programming to mitigate the issue.Purpose To analyse P:R ratios in 24-hour ECG recording using an AI-driven toolMethods We conducted a study involving 12 healthy adult volunteers and 30 adult patients with various cardiac conditions including cardiomyopathy, heart failure, and congenital heart disease. A 24-hour Holter monitor was implemented, with ECG leads positioned to replicate EVICD vectors. Initially, P, R and T waves of ECG recordings were manually annotated, followed by automated annotation using a computer algorithm. The labelled ECG data were then used to train the CNN neural network to detect P:R ratios in 10-second intervals over a 24-hour period. To ensure model robustness and prevent overfitting - where a machine learning model becomes overly tailored to the training data and fails to generalise to new data - we employed a 10-fold cross-validation strategy using the validation dataset.Results The results indicate no signs of overfitting ( figure 1), as all ten folds achieved a training and validation RMSE of less than 0.1. Analysis of the 24-hour ECG recordings revealed that nearly 60% of the study population, including healthy volunteers, exhibited periods where the P:R ratios exceeded 1:3 in one or more ECG vectors. However, 93% of the participants had at least one vector in which the P:R ratio remined below 1:3 throughout the 24-hours period.Conclusion(s) In this study, the machine learning model demonstrated excellent performance in detecting ECG waves and calculating P:R ratios over a 24-hour period. While the P:R ratio trends were observed to fluctuate throughout the day, the findings suggest that at least one vector consistently maintains a programmable P:R ratio, which could be leveraged to optimise device programming and reduce the risk of P-wave oversensing.Abstract 4-032 Figure 1Convergence of P:R ratio predictorAbstract 4-032 Figure 2An example of P:R ratio analysis over 24-hour",
  "authors": [
    {
      "affiliations": [
        "University Hospital Southampton NHS Foundation Trust, Southampton, UK",
        "University of Southampton, Southampton, UK"
      ],
      "name": "Lin-Thiri Toon"
    },
    {
      "affiliations": [
        "University Hospital Southampton NHS Foundation Trust, Southampton, UK"
      ],
      "name": "Mohamed ElRefai"
    },
    {
      "affiliations": [
        "University Hospital Southampton NHS Foundation Trust, Southampton, UK"
      ],
      "name": "Benedict M Wiles"
    },
    {
      "affiliations": [
        "University of Southampton, Southampton, UK"
      ],
      "name": "Samuel Ward"
    },
    {
      "affiliations": [
        "University of Southampton, Southampton, UK"
      ],
      "name": "Anthony J Dunn"
    },
    {
      "affiliations": [
        "University of Southampton, Southampton, UK"
      ],
      "name": "Alain B Zemkoho"
    },
    {
      "affiliations": [
        "University Hospital Southampton NHS Foundation Trust, Southampton, UK",
        "University of Southampton, Southampton, UK"
      ],
      "name": "Paul R Roberts"
    },
    {
      "affiliations": [
        "University Hospital Southampton NHS Foundation Trust, Southampton, UK",
        "University of Southampton, Southampton, UK"
      ],
      "name": "John R Paisey"
    }
  ],
  "title": "4-032 Analysis of P:R ratios over a 24-hour period: using a convoluted neural network machine learning artificial intelligence tool",
  "uid": "9813db4d-afea-5d57-ba4c-d8a3333c2fde"
}