Cumulative learning enables convolutional neural network representations for small mass spectrometry data...

Area of Science:

• Biomedical Engineering
• Computational Biology
• Data Science

Background:

• Accurate clinical diagnosis is crucial but challenging.
• Mass Spectrometry (MS) data analysis for diagnosis requires effective classification models.
• Existing Machine Learning models need time-consuming preprocessing for MS data, hindering rapid analysis.

Purpose of the Study:

• To investigate transfer learning on 1D Convolutional Neural Networks (CNNs) for MS data classification.
• To develop a cumulative learning method to enhance classification accuracy with small datasets.
• To address the limitations of traditional ML models in rapid MS data analysis.

Main Methods:

• Utilized 1D Convolutional Neural Networks (CNNs) for direct learning from raw MS data.
• Implemented transfer learning and a novel cumulative learning strategy.
• Trained models sequentially on diverse small datasets, starting with rat brain MS data.

Main Results:

• Achieved over 98% classification accuracy for 1D clinical MS data using cumulative learning.
• Demonstrated the effectiveness of cumulative learning across different biological contexts, organisms, and instruments.
• Showcased improved MS data classification performance even with limited training samples.

Conclusions:

• Cumulative learning is a promising strategy for improving MS data classification accuracy.
• This method overcomes the challenge of small sample sizes in medical data analysis.
• Enables more rapid and accurate clinical diagnosis using MS data.