The conventional clinical models predict bleeding in patients on anticoagulation therapy based on a one-time baseline measurement. Herein, we demonstrate the benefit of using time series follow-up information to improve bleeding prediction and introduce an ensemble of LSTM RNN and feedforward neural network that can use both the baseline and follow-up information to predict bleeding with AUROC of 82%, which is 14% higher than the currently-used best performing clinical model.

The performance of the clinical models used to predict bleeding in patients on anticoagulants with Venous Thromboembolism has stayed at around 70% as measured with the area under the ROC curve. In this paper, we explored the baseline information of 2542 patients using unsupervised dimensionality reduction and clustering algorithms and demonstrated a high degree of similarity between their baseline datasets. Furthermore, we observed no benefits in using supervised machine learning algorithms to predict bleeding risk compared to the conventional clinical models.

Mass Spectrometry Imaging (MSI) captures the spatial localization of the peptides and proteins on a tissue. The data obtained from MSI are analyzed using the algorithms designed for the standard tandem mass spectrometry-based proteomics; therefore, the spatial information obtained by MSI are not being used to improve protein identifications. In this thesis, we present spatially-aware protein identification algorithm for mass spectrometry imaging (SAPID-MSI) which was able to identify on average 20% more proteins at a <1% false discovery rate compared to the benchmarked algorithms.