Reported Data During Cancer Chemotherapy Course Using LSTM Modeling
By Joseph Finkelstein, Aref Smiley, Christina Echeverria, Kathi Mooney
Dept. Biomedical Informatics and College of Nursing, The University of Utah
The paper presents a study that applies Long Short-Term Memory (LSTM) modeling to predict symptom deterioration during a cancer chemotherapy course. Symptom management is crucial in cancer care to enhance patients’ quality of life by minimizing symptom severity and treatment side effects. However, guidelines for initiating symptom management strategies are not well defined. This study explores the potential of deep learning methodologies, specifically LSTM modeling, to identify serial data patterns characteristic of emerging symptom deterioration.
Data was collected from 349 patients undergoing chemotherapy, who were asked to maintain a daily symptom diary. The study focused on 12 symptoms, with each symptom assigned two values – Severity and Distress. A total symptom score was calculated for each patient, ranging from 0 to 230. After excluding patients who participated for less than three days, the final dataset contained 339 patients.
The LSTM model was implemented using MATLAB and comprised four layers. The model was trained on 80% of the data and tested on the remaining 20%. Seven LSTM models were developed to predict total scores, using symptoms from each day and the previous day. The study also incorporated a ‘cycle change’ feature to determine its impact on the model’s performance. This feature represented the changes in chemotherapy cycles for each patient.
The results showed that the LSTM model could predict symptom scores for up to a week. When the ‘cycle change’ feature was included, the R-squared values indicated a reasonable level of prediction accuracy. However, the Root Mean Square Error (RMSE) values suggested room for improvement in the model’s predictions. Excluding the ‘cycle change’ feature resulted in similar R-squared and RMSE values, suggesting that its inclusion did not significantly improve the model’s performance.
In conclusion, the study demonstrates the potential of LSTM models in predicting cancer symptom progression, offering valuable insights for personalized symptom management in cancer patients. Future research will focus on balancing the dataset and refining the deep learning algorithms to enhance the model’s predictive capabilities. This study represents an important step forward in the integration of deep learning methodologies into cancer symptom management, with the potential to significantly improve patient care.
