Functional principal component analysis for longitudinal data with informative dropout

Evaluating Prognostic Value of Dynamics of Circulating Lactate Dehydrogenase in Colorectal Cancer Using Modeling and Machine Learning

Pretreatment serum lactate dehydrogenase (LDH) levels have been associated with poor prognosis in several types of cancer, including metastatic colorectal cancer (mCRC). However, very few models link survival to longitudinal LDH measured repeatedly over time during treatment. We investigated the prognostic value of on-treatment LDH dynamics in mCRC. Using data from two large phase III studies (2L and 3L+ mCRC settings, n = 824 and 210, respectively), we found that integrating longitudinal LDH data with baseline risk factors significantly improved survival prediction. Current LDH values performed best, enhancing discrimination ability (area under the receiver operating characteristic curve) by 4.5~15.4% and prediction accuracy (Brier score) by 3.9~15.0% compared with baseline variables. Combining all longitudinal LDH markers further improved predictive performance. After controlling for baseline covariates and other longitudinal LDH indicators, current LDH levels remained a significant risk factor in mCRC, increasing mortality risk by over 90% (P < 0.001) in 2L patients and 60-70% (P < 0.01) in 3L+ patients per unit increment in current log (LDH). Machine-learning techniques, like functional principal component analysis (FPCA), extracted informative features from longitudinal LDH data, capturing over 99% of variability and allowing prediction of survival. Unsupervised clustering based on the extracted FPCA features stratified patients into three groups with distinct LDH dynamics and survival outcomes. Hence, our approaches offer a valuable and cost-effective way for risk stratification and improves survival prediction in mCRC using LDH trajectories.

Pooling random forest and functional data analysis for biomedical signals supervised classification: Theory and application to electrocardiogram data

Scientific progress has contributed to creating many devices to gather vast amounts of biomedical data over time. The goal of these devices is generally to monitor people's health conditions, diagnose, and prevent patients' diseases, for example, to discover cardiovascular disorders or predict epileptic seizures. A common way of investigating these data is classification, but these instruments generate signals often characterized by high dimensionality. Learning from these data is definitely a challenging task due to many issues, for example, the trade‐off between complexity and accuracy and the course of dimensionality. This study proposes a supervised classification method based on the joint use of functional data analysis, classification trees, and random forest to deal with massive biomedical data recorded over time. For this purpose, this research suggests different original tools to extract features and train functional classifiers, interpret the classification rules, assess leaves' quality and composition, avoid the classical drawbacks due to the COD, and improve the accuracy of the functional classifiers. Focusing on ECG data as a possible example, the final purpose of this study is to offer an original approach to identify and classify patients at risk using different types of biomedical signals. The results confirm that this line of research is exciting; indeed, the interpretative tools show evidence to be very useful for understanding classification rules. Furthermore, the performance of the proposed functional classifier, in terms of accuracy, is excellent because the latter breaks the previous classification record regarding a well‐known ECG dataset.