Interpreting Mass Spectrometry Images Using Stacked Learning
DESI , Mass Spectrometry , Deep Learning , Stacked Learning , Atrial Fibrillation
Desorption electrospray ionization mass spectrometry (DESI-MS) allows for hyperspectral analysis of tissue samples non-destructively under ambient conditions. A downside of such a bountiful source of data is that the high dimensional images suffer from the curse of dimensionality, often requiring substantial pre-processing and complex algorithmic approaches to extract meaningful interpretations of the data. Postoperative atrial fibrillation (POAF) is a cardiac arrhythmia that can occur after heart surgery, resulting in poorer recovery and higher treatment costs. We proposed a stacked learning system using a combination of DESI-MS and patient demographic information to classify POAF in a set of patients. This stacked learning system was composed of three stages. First, a CNN was trained to identify nuclear hypertrophy in DESI-MS images. Second, the CNN’s output was combined with patient demographic information and dimensionally reduced. Finally, a support vector machine was used to classify POAF on the resulting dataset. Our study showed that the full stacked learning system outperformed the individual components and also provided a preliminary step toward predicting POAF in patients.