[Prognostic model of small sample critical diseases based on transfer learning]

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Prognostics

Background:

• Small sample sizes in critical disease prognosis lead to overfitting, high prediction error, and instability in models.
• Traditional prognostic models struggle with limited clinical data, hindering accurate patient outcome prediction.

Purpose of the Study:

• To introduce the long short-term memory transferring algorithm (transLSTM) for improving prognostic models with small critical disease samples.
• To leverage transfer learning to enhance prediction performance and reduce data requirements for critical disease prognosis.

Main Methods:

• The transLSTM algorithm utilizes transfer learning by pretraining model parameters on related disease data before fine-tuning with target disease samples.
• It transfers information between disease prognostic models to construct effective models for small-sample target diseases.

Main Results:

• On the MIMIC-Ⅲ database, transLSTM demonstrated higher AUROC (0.02-0.07) and AUPRC (0.05-0.14) compared to traditional LSTM.
• TransLSTM required 39%-64% fewer training iterations and achieved comparable sepsis mortality prediction with 100 samples versus 250 for traditional models.

Conclusions:

• The transLSTM algorithm significantly improves prediction accuracy and training speed for critical disease prognostic models using small sample sizes.
• This study successfully applies transfer learning to enhance prognostic modeling in scenarios with limited clinical data.