As part of the PhysioNet 2023 Challenge, our team FMMGroup_UVa, presents an original approach for the prediction of the outcome of patients after a heart attack.
Our approach considers a 10-component decomposition of ten seconds 18-dimensional EEG epochs using the 3DFMM model. The components describe waves that are common to all channels and are parametrized in terms of 2 parameters describing the common-to-all wave location and sharpness, and 2x10 parameters describing the single-channel wave amplitudes and symmetries. We create features using statistics from these parameters that capture EEG core aspects for the prognosis, such as amplitude, complexity, connectivity, regularity, or spike counting. We developed a prediction index for each epoch using binary classifiers defined when selected features are out of normal ranges, which are derived from the training epochs of patients with good prognoses. The patient average epoch indexes, the shockable rhythm, and the age are included in a classification model to obtain an outcome score in a given time block.
Our preliminary proposal considers 28 features, OneR binary classifiers at 12H and 24H, and boosting binary classifiers for predictions at 48H and 72H. We have obtained, for a specificity of 0.95, sensitivity values of: 0.61 at 12H, 0.63 at 24H, 0.62 at 48H, and 0.58 at 72H, using 5-fold cross-validation. Using 10% of the training data, the unofficial phase score is 0.33 at 72H (a single entry rank 79th out of 159). We plan to improve the efficiency of our program and to increase by a large extent our scores in the challenge, using the entire data, adding new features, and further investigating additional machine learning techniques. The interpretability of the features and the prediction rules, and the relatively good performance at 12H and 24H are the main strengths of our method.