{
  "abstract": "Introduction Interpretation of CPET is complex involving both tabular and graphical data analysis. Recently machine learning techniques have been used to improve CPET interpretation accuracy (Portella et al. IEEE J. Biomed. Health Inform 2022;26:4228-4237). Despite the potential for these techniques, only specific patient cohorts have been examined so far. The aim of this study was to use machine learning techniques to identify associations between CPET variables and primary diagnosis.Methods Retrospective, harmonised, multi-source datasets were used (N=320) to explore associations between clinical predictors and CPET variables. Bivariate comparisons used Fisher’s exact test and Mann–Whitney U tests. Multivariable models were fitted to adjust for age and sex: logistic regression and ordinary least squares regression. FDR adjusted statistical significance was defined as q<0.05.Supervised machine learning techniques including Elastic Net, Linear SVM, and Random Forest Classifier were used to identify distinct physiological phenotypes associated with Vascular, Cardiac, Respiratory and Normal diagnosis using CPET data. Nested cross-validation framework was used to estimate model performance.Results The Respiratory diagnosis group was associated with lower FEV1 (−0.50 L; q=6×10 −6; N=311), lower workload (−19.32 W; q=0.023; N=320) and lower peak SpO2 (SpO2 peak −3.86%; q=1.5×10−5; N=198).The Cardiac diagnosis group was associated with a lower peak VO2 (−343.55 mL/min; q=0.00057; N=320), lower peak O2 pulse (−2.04 mL; q=0.00088; N=320), lower peak VE (−10.73 L/min; q=0.0088; N=320), and higher resting systolic BP (+8.58 mmHg; q=0.042; N=277).The Vascular diagnosis group was associated with a higher VE/VCO2 peak (+12.66; q=7.6×10−12; N=320), lower peak VO2 (−674.88 mL/min; q=2.1×10−6;N=320), and a lower peak O2 pulse (−3.95 mL; q=6.4×10−6; N=320).The Vascular Model identified VE/VCO2 (rest) as the strongest predictor. The Respiratory Model identified low FEV1, high VE/VCO2 (peak), low PETCO2 (rest), low SpO2 (peak) and low O2 pulse. The Cardiac Model identified a low VO2 (peak), high VE/VCO2 (peak) and PETO2 (rest). Table 1Conclusion This work provides an insight into how machine learning techniques can be used to identify predictions amongst CPET variables and primary diagnosis. Future work will examine the use of algorithm based models which utilise these key diagnostic criteria in the conditions highlighted.Abstract P8 Table 1Machine learning algorithmsDiagnosisAlgorithmNPrevalenceCV_Bal_AccCV_AUCCardiacElasticNet13538.5%0.777 ± 0.0800.870 ± 0.071CardiacLinearSVM13538.5%0.772 ± 0.0600.869 ± 0.066NormalElasticNet32026.2%0.726 ± 0.0470.797 ± 0.045NormalLinearSVM32026.2%0.724 ± 0.0450.803 ± 0.039RespiratoryLinearSVM18055.0%0.705 ± 0.0680.775 ± 0.058RespiratoryElasticNet18055.0%0.698 ± 0.0690.754 ± 0.068VascularLinearSVM9410.6%0.848 ± 0.1260.927 ± 0.086VascularElasticNet9410.6%0.845 ± 0.1730.951 ± 0.070",
  "authors": [
    {
      "affiliations": [
        "Royal Berkshire NHS Foundation Trust, Reading, United Kingdom"
      ],
      "name": "Jessica Swan"
    },
    {
      "affiliations": [
        "Royal Berkshire NHS Foundation Trust, Reading, United Kingdom"
      ],
      "name": "Mark Unstead"
    },
    {
      "affiliations": [
        "Manchester Metropolitan University, Manchester, United Kingdom"
      ],
      "name": "Liam Bagley"
    },
    {
      "affiliations": [
        "Intelligify Limited, London, United Kingdom"
      ],
      "name": "Alla Veselkova"
    },
    {
      "affiliations": [
        "Intelligify Limited, London, United Kingdom"
      ],
      "name": "Ivan Laponogov"
    },
    {
      "affiliations": [
        "University Hospitals of Leicester NHS Trust, Leicester, United Kingdom"
      ],
      "name": "Natalie Wilson"
    },
    {
      "affiliations": [
        "Royal Berkshire NHS Foundation Trust, Reading, United Kingdom"
      ],
      "name": "Catherine Thomas"
    }
  ],
  "title": "P8 Use of machine learning in cardiopulmonary exercise testing to identify diagnosis outcomes",
  "uid": "2af2b981-7189-5c6a-9d7c-419f51c6eaf1"
}