ICU Outcome Prediction

PhysioNet 2012 Dataset and the mTAND Architecture

  • The PhysioNet dataset challenges ML systems in predicting patient outcomes in the ICU

  • Only 4000 datapoints

  • Up to 42 variables 

    • Six of these variables are general descriptors (collected on admission)

    • Remainder are time series, for which multiple observations may be available.

  • Multi-Time Attention Networks for Irregularly Sampled Time Series (mTAND)

    • State-of-the-art open-source architecture on the PhysioNet Dataset at time of test

Diagram illustrating the mTAND model, which processes irregularly sampled multivariate time series using attention mechanisms, reference points, and a linear layer to produce outputs based on reference points.
  • Perforated Backpropagation overfit when added to original architecture

    • We then increased and decreased the width of the layers of the network by a constant factor X, giving “Net X” in the graph

      • No other modifications were made to depth or any other parameters in the training pipeline

  • Graph shows Test AUC scores which is the metric used on the PhysioNet competition leaderboard, along with parameter counts of each network

  • Doubling the network width has very small effect on accuracy

  • Reducing width and adding Perforated AI technology to the reduced network resulted in a final network size which was 9.3% the initial size while improving accuracy by 4%!

    • Each point shows the AUC score as additional Dendrites are added

Line graph showing mTAN on PhysioNet results with AUC scores on y-axis and parameters in millions on log scale on x-axis. Different colored lines represent Width*0.125, Width*0.25, Width*0.5, Original, and Width*2.