Prediction of Cancer Symptom Trajectory Using Longitudinal Electronic Health Record Data and Long Short-Term Memory...

Area of Science:

• Oncology
• Medical Informatics
• Computational Biology

Background:

• Accurate prediction of cancer symptom severity and progression is crucial for timely clinical intervention and treatment planning.
• Current prediction methods are often limited by sparse, inconsistent data and simplistic measures like last observed symptom severity.
• Developing robust predictive models is essential for improving patient outcomes and quality of life in cancer care.

Purpose of the Study:

• To develop and evaluate a predictive model for future cancer symptom experiences based on historical symptom data.
• To compare the performance of different machine learning models in predicting symptom trajectories.
• To leverage routinely collected nursing documentation for enhanced cancer symptom prediction.

Main Methods:

• Retrospective, longitudinal analysis of 208 hospitalized cancer patients' records (2008-2014).
• Training and evaluation of Long Short-Term Memory (LSTM) recurrent neural networks, linear regression, and random forest models.
• Models were trained on past symptom data to predict future symptom trajectories.

Main Results:

• At least one tested model (LSTM, linear regression, random forest) surpassed predictions based solely on previous clinical observation.
• LSTM models significantly outperformed linear regression and random forest in predicting nausea and psychosocial status.
• Linear regression excelled in predicting oral health, while random forest was superior for mobility and nutrition predictions.

Conclusions:

• Routinely collected nursing documentation, even with sparse data, can be used to successfully predict patient symptom trajectories.
• The developed prediction models can aid in individualizing symptom management strategies for cancer patients.
• Improved symptom prediction has the potential to significantly support cancer patients' quality of life.