Dynamic and concordance-assisted learning for risk stratification with application to Alzheimer's disease.

Dynamic prediction models capable of retaining accuracy by evolving over time could play a significant role for monitoring disease progression in clinical practice. In biomedical studies with long-term follow up, participants are often monitored through periodic clinical visits with repeat measurements until an occurrence of the event of interest (e.g. disease onset) or the study end. Acknowledging the dynamic nature of disease risk and clinical information contained in the longitudinal markers, we propose an innovative concordance-assisted learning algorithm to derive a real-time risk stratification score. The proposed approach bypasses the need to fit regression models, such as joint models of the longitudinal markers and time-to-event outcome, and hence enjoys the desirable property of model robustness. Simulation studies confirmed that the proposed method has satisfactory performance in dynamically monitoring the risk of developing disease and differentiating high-risk and low-risk population over time. We apply the proposed method to the Alzheimer's Disease Neuroimaging Initiative data and develop a dynamic risk score of Alzheimer's Disease for patients with mild cognitive impairment using multiple longitudinal markers and baseline prognostic factors.

Deep learning models to predict primary open-angle glaucoma.

Glaucoma is a major cause of blindness and vision impairment worldwide, and visual field (VF) tests are essential for monitoring the conversion of glaucoma. While previous studies have primarily focused on using VF data at a single time point for glaucoma prediction, there has been limited exploration of longitudinal trajectories. Additionally, many deep learning techniques treat the time-to-glaucoma prediction as a binary classification problem (glaucoma Yes/No), resulting in the misclassification of some censored subjects into the nonglaucoma category and decreased power. To tackle these challenges, we propose and implement several deep-learning approaches that naturally incorporate temporal and spatial information from longitudinal VF data to predict time-to-glaucoma. When evaluated on the Ocular Hypertension Treatment Study (OHTS) dataset, our proposed convolutional neural network (CNN)-long short-term memory (LSTM) emerged as the top-performing model among all those examined. The implementation code can be found online (https://github.com/rivenzhou/VF_prediction).

Dynamic prediction of lung cancer suicide risk based on meteorological factors and clinical characteristics：A landmarking analysis approach.

Suicide is a severe public health issue globally. Accurately identifying high-risk lung cancer patients for suicidal behavior and taking timely intervention measures has become a focus of current research. This study intended to construct dynamic prediction models for identifying suicide risk among lung cancer patients. Patients were sourced from the Surveillance, Epidemiology, and End Results database, while meteorological data was acquired from the Centers for Disease Control and Prevention. This cohort comprised 455, 708 eligible lung cancer patients from January 1979 to December 2011. A Cox proportional hazard regression model based on landmarking approach was employed to explore the impact of meteorological factors and clinical characteristics on suicide among lung cancer patients, and to build dynamic prediction models for the suicide risk of these patients. Additionally, subgroup analyses were conducted by age and sex. The model's performance was evaluated using the C-index, Brier score, area under curve (AUC) and calibration plot. During the study period, there were 666 deaths by suicide among lung cancer patients. Multivariable Cox results from the dynamic prediction model indicated that age, marital status, race, sex, primary site, stage, monthly average daily sunlight, and monthly average temperature were significant predictors of suicide. The dynamic prediction model demonstrated well consistency and discrimination capabilities. Subgroup analyses revealed that the association of monthly average daily sunlight and monthly average temperature with suicide remained significant among female and younger lung cancer patients. The dynamic prediction model can effectively incorporate covariates with time-varying to predict lung cancer patients' suicide death. The results of this study have significant implications for assessing lung cancer individuals' suicide risk.

Delayed kernels for longitudinal survival analysis and dynamic prediction.

Predicting patient survival probabilities based on observed covariates is an important assessment in clinical practice. These patient-specific covariates are often measured over multiple follow-up appointments. It is then of interest to predict survival based on the history of these longitudinal measurements, and to update predictions as more observations become available. The standard approaches to these so-called 'dynamic prediction' assessments are joint models and landmark analysis. Joint models involve high-dimensional parameterizations, and their computational complexity often prohibits including multiple longitudinal covariates. Landmark analysis is simpler, but discards a proportion of the available data at each 'landmark time'. In this work, we propose a 'delayed kernel' approach to dynamic prediction that sits somewhere in between the two standard methods in terms of complexity. By conditioning hazard rates directly on the covariate measurements over the observation time frame, we define a model that takes into account the full history of covariate measurements but is more practical and parsimonious than joint modelling. Time-dependent association kernels describe the impact of covariate changes at earlier times on the patient's hazard rate at later times. Under the constraints that our model (a) reduces to the standard Cox model for time-independent covariates, and (b) contains the instantaneous Cox model as a special case, we derive two natural kernel parameterizations. Upon application to three clinical data sets, we find that the predictive accuracy of the delayed kernel approach is comparable to that of the two existing standard methods.

Dynamic Relapse Prediction by Peripheral Blood WT1mRNA after Allogeneic Hematopoietic Cell Transplantation for Myeloid Neoplasms.

Although various relapse prediction models based on pretransplant information have been reported, they cannot update the predictive probability considering post-transplant patient status. Therefore, these models are not appropriate for deciding on treatment adjustment and preemptive intervention during post-transplant follow-up. A dynamic prediction model can update the predictive probability by considering the information obtained during follow-up. This study aimed to develop and assess a dynamic relapse prediction model after allogeneic hematopoietic cell transplantation (allo-HCT) for acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS) using peripheral blood Wilms' tumor 1 messenger RNA (WT1mRNA). We retrospectively analyzed patients with AML or MDS who underwent allo-HCT at our institution. To develop dynamic models, we employed the landmarking supermodel approach, using age, refined disease risk index, conditioning intensity, and number of transplantations as pretransplant covariates and both pre- and post-transplant peripheral blood WT1mRNA levels as time-dependent covariates. Finally, we compared the predictive performances of the conventional and dynamic models by area under the time-dependent receiver operating characteristic curves. A total of 238 allo-HCT cases were included in this study. The dynamic model that considered all pretransplant WT1mRNA levels and their kinetics showed superior predictive performance compared to models that considered only pretransplant covariates or factored in both pretransplant covariates and post-transplant WT1mRNA levels without their kinetics; their time-dependent areas under the curve were 0.89, 0.73, and 0.87, respectively. The predictive probability of relapse increased gradually from approximately 90 days before relapse. Furthermore, we developed a web application to make our model user-friendly. This model facilitates real-time, highly accurate, and personalized relapse prediction at any time point after allo-HCT. This will aid decision-making during post-transplant follow-up by offering objective relapse forecasts for physicians.

Dynamic Prediction of Overall Survival for Patients with Osteosarcoma: A Retrospective Analysis of the EURAMOS-1 Clinical Trial Data.

Current prediction models for patients with ostosarcoma are restricted to predictions from a single, static point in time, such as diagnosis or surgery. These approaches discard information which becomes available during follow-up and may have an impact on patient's prognosis. This study aims at developing a dynamic prediction model providing 5-year overall survival (OS) predictions from different time points during follow-up. The developed model considers relevant baseline prognostic factors, accounting for where appropriate time-varying effects and time-varying intermediate events such as local recurrence (LR) and new metastatic disease (NM). A landmarking approach is applied to 1965 patients with high-grade resectable osteosarcoma from the EURAMOS-1 trial (NCT00143030). Results show that LR and NM negatively affected 5-year OS (HRs: 2.634, 95% CI 1.845-3.761; 8.558, 95% CI 7.367-9.942, respectively). Baseline factors with strong prognostic value (HRs > 2) included poor histological response (≥10% viable tumor), axial tumor location, and the presence of lung metastases. The effect of poor versus good histological response changed over time, becoming non-significant from 3.25 years post-surgery onwards. This time-varying effect, as well as the strong impact of disease-related time-varying variables, show the importance of including updated information collected during follow-up in the model to provide more accurate survival predictions.

A dynamic prediction model for preeclampsia using the sFlt-1/PLGF ratio combined with multiple factors.

OBJECTIVE: Preeclampsia (PE) is a pregnancy-related multi-organ disease and a significant cause of incidence rate and mortality of pregnant women and newborns worldwide. Delivery remains the only available treatment for PE. This study aims to establish a dynamic prediction model for PE. METHODS: A total of 737 patients who visited our hospital from January 2021 to June 2022 were identified according to the inclusion and exclusion criteria, forming the primary dataset. Additionally, 176 singleton pregnant women who visited our hospital from July 2022 to November 2022 comprised the verification set. We investigated different gestational weeks of sFlt-1/PLGF (soluble FMS-like tyrosine kinase-1, placental growth factor) ratio combined with maternal characteristics and routine prenatal laboratory results in order to predict PE in each trimester. Multivariate logistic regression was used to establish the prediction model for PE at different gestational weeks. The discrimination, calibration, and clinical validity were utilized to evaluate predictive models as well as models in external validation queues. RESULTS: At 20-24 weeks, the obtained prediction model for PE yielded an area under the curve of 0.568 (95% confidence interval, 0.479-0.657). At 25-29 weeks, the obtained prediction model for PE yielded an area under the curve of 0.773 (95% confidence interval, 0.703-0.842)and 0.731 (95% confidence interval, 0.653-0.809) at 30-34 weeks. After adding maternal factors, uterine artery pulsation index(Ut-IP), and other laboratory indicators to the sFlt-1/PLGF ratio, the predicted performance of PE improved. It found that the AUC improved to 0.826(95% confidence interval, 0.748 ∼ 0.904) at 20-24 weeks, 0.879 (95% confidence interval, 0.823 ∼ 0.935) at 25-29 weeks, and 0.862(95% confidence interval, 0.799 ∼ 0.925) at 30-34 weeks.The calibration plot of the prediction model indicates good predictive accuracy between the predicted probability of PE and the observed probability. Furthermore, decision-curve analysis showed an excellent clinical application value of the models. CONCLUSION: Using the sFlt-1/PLGF ratio combined with multiple factors at 25-29 weeks can effectively predict PE, but the significance of re-examination in late pregnancy is not significant.

Serial platelet count as a dynamic prediction marker of hospital mortality among septic patients.

Background: Platelets play a critical role in hemostasis and inflammatory diseases. Low platelet count and activity have been reported to be associated with unfavorable prognosis. This study aims to explore the relationship between dynamics in platelet count and in-hospital morality among septic patients and to provide real-time updates on mortality risk to achieve dynamic prediction. Methods: We conducted a multi-cohort, retrospective, observational study that encompasses data on septic patients in the eICU Collaborative Research Database (eICU-CRD) and the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. The joint latent class model (JLCM) was utilized to identify heterogenous platelet count trajectories over time among septic patients. We assessed the association between different trajectory patterns and 28-day in-hospital mortality using a piecewise Cox hazard model within each trajectory. We evaluated the performance of our dynamic prediction model through area under the receiver operating characteristic curve, concordance index (C-index), accuracy, sensitivity, and specificity calculated at predefined time points. Results: Four subgroups of platelet count trajectories were identified that correspond to distinct in-hospital mortality risk. Including platelet count did not significantly enhance prediction accuracy at early stages (day 1 C-indexDynamic  vs C-indexWeibull: 0.713 vs 0.714). However, our model showed superior performance to the static survival model over time (day 14 C-indexDynamic  vs C-indexWeibull: 0.644 vs 0.617). Conclusions: For septic patients in an intensive care unit, the rapid decline in platelet counts is a critical prognostic factor, and serial platelet measures are associated with prognosis.

Research on Intelligent Vehicle Trajectory Tracking Control Based on Improved Adaptive MPC.

Intelligent vehicle trajectory tracking exhibits problems such as low adaptability, low tracking accuracy, and poor robustness in complex driving environments with uncertain road conditions. Therefore, an improved method of adaptive model predictive control (AMPC) for trajectory tracking was designed in this study to increase the corresponding tracking accuracy and driving stability of intelligent vehicles under uncertain and complex working conditions. First, based on the unscented Kalman filter, longitudinal speed, yaw speed, and lateral acceleration were considered as the observed variables of the measurement equation to estimate the lateral force of the front and rear tires accurately in real time. Subsequently, an adaptive correction estimation strategy for tire cornering stiffness was designed, an AMPC method was established, and a dynamic prediction time-domain adaptive model was constructed for optimization according to vehicle speed and road adhesion conditions. The improved AMPC method for trajectory tracking was then realized. Finally, the control effectiveness and trajectory tracking accuracy of the proposed AMPC technique were verified via co-simulation using CarSim and MATLAB/Simulink. From the results, a low lateral position error and heading angle error in trajectory tracking were obtained under different vehicle driving conditions and road adhesion conditions, producing high trajectory-tracking control accuracy. Thus, this work provides an important reference for improving the adaptability, robustness, and optimization of intelligent vehicle tracking control systems.

Two-step interpretable modeling of ICU-AIs.

We present a novel methodology for integrating high resolution longitudinal data with the dynamic prediction capabilities of survival models. The aim is two-fold: to improve the predictive power while maintaining the interpretability of the models. To go beyond the black box paradigm of artificial neural networks, we propose a parsimonious and robust semi-parametric approach (i.e., a landmarking competing risks model) that combines routinely collected low-resolution data with predictive features extracted from a convolutional neural network, that was trained on high resolution time-dependent information. We then use saliency maps to analyze and explain the extra predictive power of this model. To illustrate our methodology, we focus on healthcare-associated infections in patients admitted to an intensive care unit.

Utility of multimodal longitudinal imaging data for dynamic prediction of cardiovascular and renal disease: the CARDIA study.

Background: Medical examinations contain repeatedly measured data from multiple visits, including imaging variables collected from different modalities. However, the utility of such data for the prediction of time-to-event is unknown, and only a fraction of the data is typically used for risk prediction. We hypothesized that multimodal longitudinal imaging data could improve dynamic disease prognosis of cardiovascular and renal disease (CVRD). Methods: In a multi-centered cohort of 5,114 CARDIA participants, we included 166 longitudinal imaging variables from five imaging modalities: Echocardiography (Echo), Cardiac and Abdominal Computed Tomography (CT), Dual-Energy x-ray Absorptiometry (DEXA), Brain Magnetic Resonance Imaging (MRI) collected from young adulthood to mid-life over 30 years (1985-2016) to perform dynamic survival analysis of CVRD events using machine learning dynamic survival analysis (Dynamic-DeepHit, LTRCforest, and Extended Cox for Time-varying Covariates). Risk probabilities were continuously updated as new data were collected. Model performance was assessed using integrated AUC and C-index and compared to traditional risk factors. Results: Longitudinal imaging data, even when being irregularly collected with high missing rates, improved CVRD dynamic prediction (0.03 in integrated AUC, up to 0.05 in C-index compared to traditional risk factors; best model's C-index = 0.80-0.83 up to 20 years from baseline) from young adulthood followed up to midlife. Among imaging variables, Echo and CT variables contributed significantly to improved risk estimation. Echo measured in early adulthood predicted midlife CVRD risks almost as well as Echo measured 10-15 years later (0.01 C-index difference). The most recent CT exam provided the most accurate prediction for short-term risk estimation. Brain MRI markers provided additional information from cardiac Echo and CT variables that led to a slightly improved prediction. Conclusions: Longitudinal multimodal imaging data readily collected from follow-up exams can improve CVRD dynamic prediction. Echocardiography measured early can provide a good long-term risk estimation, while CT/calcium scoring variables carry atherosclerotic signatures that benefit more immediate risk assessment starting in middle-age.

Blood biomarkers in dynamic prediction of conversion to Alzheimer's disease: An application of joint modeling.

OBJECTIVES: To investigate the accuracy of longitudinal trajectories of blood biomarkers for predicting future onset of AD among MCI participants as well as to demonstrate dynamic prediction of the individual conversion risk applying joint modeling. METHODS: A total of 446 participants with MCI at baseline from the Alzheimer's Disease Neuroimaging Initiative database were included. We introduced joint modeling to analyze the effects of the longitudinal blood biomarkers on the conversion risk to AD, and further to build individual-specific prediction risk model. RESULTS: During the follow-up, 345 participants remained with MCI and 101 progressed to AD, and were categorized as non-progression and progression group, respectively. Longitudinally, the positive association of the concentration dynamics of plasma p-tau181 and NfL with the conversion risk to AD from MCI was also demonstrated, with Hazard Ratio (HR) = 5.83 and HR = 4.18, respectively. When incorporating plasma p-tau181 and NfL together to predict AD progression, we observed improved performance (AUC = 0.701, Brier Score = 0.119). Two participants were chosen to exemplify the individual-specific risk prediction at different follow-up time for comparative analysis. CONCLUSIONS: Plasma p-tau181 and NfL could serve as biomarkers for the prediction of AD onset, and the individualized prediction opens up the possibility to provide clinical information at a personal level.

Dynamic prediction of survival using multivariate functional principal component analysis: A strict landmarking approach.

Dynamically predicting patient survival probabilities using longitudinal measurements has become of great importance with routine data collection becoming more common. Many existing models utilize a multi-step landmarking approach for this problem, mostly due to its ease of use and versatility but unfortunately most fail to do so appropriately. In this article we make use of multivariate functional principal component analysis to summarize the available longitudinal information, and employ a Cox proportional hazards model for prediction. Additionally, we consider a centred functional principal component analysis procedure in an attempt to remove the natural variation incurred by the difference in age of the considered subjects. We formalize the difference between a 'relaxed' landmarking approach where only validation data is landmarked and a 'strict' landmarking approach where both the training and validation data are landmarked. We show that a relaxed landmarking approach fails to effectively use the information contained in the longitudinal outcomes, thereby producing substantially worse prediction accuracy than a strict landmarking approach.

Dynamic Prognosis Prediction for Patients on DAPT After Drug-Eluting Stent Implantation: Model Development and Validation.

BACKGROUND: The rapid evolution of artificial intelligence (AI) in conjunction with recent updates in dual antiplatelet therapy (DAPT) management guidelines emphasizes the necessity for innovative models to predict ischemic or bleeding events after drug-eluting stent implantation. Leveraging AI for dynamic prediction has the potential to revolutionize risk stratification and provide personalized decision support for DAPT management. METHODS AND RESULTS: We developed and validated a new AI-based pipeline using retrospective data of drug-eluting stent-treated patients, sourced from the Cerner Health Facts data set (n=98 236) and Optum's de-identified Clinformatics Data Mart Database (n=9978). The 36 months following drug-eluting stent implantation were designated as our primary forecasting interval, further segmented into 6 sequential prediction windows. We evaluated 5 distinct AI algorithms for their precision in predicting ischemic and bleeding risks. Model discriminative accuracy was assessed using the area under the receiver operating characteristic curve, among other metrics. The weighted light gradient boosting machine stood out as the preeminent model, thus earning its place as our AI-DAPT model. The AI-DAPT demonstrated peak accuracy in the 30 to 36 months window, charting an area under the receiver operating characteristic curve of 90% [95% CI, 88%-92%] for ischemia and 84% [95% CI, 82%-87%] for bleeding predictions. CONCLUSIONS: Our AI-DAPT excels in formulating iterative, refined dynamic predictions by assimilating ongoing updates from patients' clinical profiles, holding value as a novel smart clinical tool to facilitate optimal DAPT duration management with high accuracy and adaptability.

Online interpretable dynamic prediction models for postoperative delirium after cardiac surgery under cardiopulmonary bypass developed based on machine learning algorithms: A retrospective cohort study.

OBJECTIVE: Postoperative delirium (POD) is strongly associated with poor early and long-term prognosis in cardiac surgery patients with cardiopulmonary bypass (CPB). This study aimed to develop dynamic prediction models for POD after cardiac surgery under CPB using machine learning (ML) algorithms. METHODS: From July 2021 to June 2022, clinical data were collected from patients undergoing cardiac surgery under CPB at Nanjing First Hospital. A dataset from the same center (October 2022 to November 2022) was also used for temporal external validation. We used ML and deep learning to build models in the training set, optimized parameters in the test set, and finally validated the best model in the validation set. The SHapley Additive exPlanations (SHAP) method was introduced to explain the best models. RESULTS: Of the 885 patients enrolled, 221 (25.0%) developed POD. 22 (22.0%) of 100 validation cohort patients developed POD. The preoperative and postoperative artificial neural network (ANN) models exhibited optimal performance. The validation results demonstrated satisfactory predictive performance of the ANN model, with area under the receiver operator characteristic curve (AUROC) values of 0.776 and 0.684 for the preoperative and postoperative models, respectively. Based on the ANN algorithm, we constructed dynamic, highly accurate, and interpretable web risk calculators for POD. CONCLUSIONS: We successfully developed online interpretable dynamic ANN models as clinical decision aids to identify patients at high risk of POD before and after cardiac surgery to facilitate early intervention or care.

Research on landslide susceptibility prediction model based on LSTM-RF-MDBN.

The occurrence of landslide disasters causes huge economic losses and casualties. Although many achievements have been made in predicting the probability of landslide disasters, various factors such as the scale and spatial location of landslide geological disasters should still be fully considered. Further research on how to quantitatively characterize the susceptibility of landslide geological disasters is necessarily important. To this end, taking the Wenchuan earthquake as the research area and extracting eight influencing factors, including terrain information entropy (Ht), lithology, distance from rivers, distance from faults, vegetation coverage (NDVI), distance from roads, peak ground motion acceleration (PGA), and annual rainfall, a landslide susceptibility prediction model was hereby established based on LSTM-RF-MDBN, a landslide susceptibility prediction map was drawn, and the spatial distribution characteristics of landslide disasters were analyzed. The results showed that (1) LSTM had good prediction results for the eight influencing factors, with an average prediction accuracy of 85%; (2) compared with models such as DNN and LR for predicting landslide disaster points, the AUC value of RF for predicting landslide point positions reached 0.88, presenting a higher accuracy compared to other models; (3) the AUC value of the landslide susceptibility prediction model based on LSTM-RF-MDBN reached 0.965, which had a high accuracy in predicting landslide susceptibility. Overall, the research results can provide a scientific basis for selecting the best strategy for landslide disaster warning, prevention, and mitigation.

Illustration of association between change in prostate-specific antigen (PSA) values and time to tumor status after treatment for prostate cancer patients: a joint modelling approach.

BACKGROUND: Prostate cancer (PCa) is the most prevalent tumor in men, and Prostate-Specific Antigen (PSA) serves as the primary marker for diagnosis, recurrence, and disease-free status. PSA levels post-treatment guide physicians in gauging disease progression and tumor status (low or high). Clinical follow-up relies on monitoring PSA over time, forming the basis for dynamic prediction. Our study proposes a joint model of longitudinal PSA and time to tumor shrinkage, incorporating baseline variables. The research aims to assess tumor status post-treatment for dynamic prediction, utilizing joint assessment of PSA measurements and time to tumor status. METHODS: We propose a joint model for longitudinal PSA and time to tumor shrinkage, taking into account baseline BMI and post-treatment factors, including external beam radiation therapy (EBRT), androgen deprivation therapy (ADT), prostatectomy, and various combinations of these interventions. The model employs a mixed-effect sub-model for longitudinal PSA and an event time sub-model for tumor shrinkage. RESULTS: Results emphasize the significance of baseline factors in understanding the relationship between PSA trajectories and tumor status. Patients with low tumor status consistently exhibit low PSA values, decreasing exponentially within one month post-treatment. The correlation between PSA levels and tumor shrinkage is evident, with the considered factors proving to be significant in both sub-models. CONCLUSIONS: Compared to other treatment options, ADT is the most effective in achieving a low tumor status, as evidenced by a decrease in PSA levels after months of treatment. Patients with an increased BMI were more likely to attain a low tumor status. The research enhances dynamic prediction for PCa patients, utilizing joint analysis of PSA and time to tumor shrinkage post-treatment. The developed model facilitates more effective and personalized decision-making in PCa care.

DYNAMIC PREDICTION OF RESIDUAL LIFE WITH LONGITUDINAL COVARIATES USING LONG SHORT-TERM MEMORY NETWORKS.

Sepsis, a complex medical condition that involves severe infections with life-threatening organ dysfunction, is a leading cause of death worldwide. Treatment of sepsis is highly challenging. When making treatment decisions, clinicians and patients desire accurate predictions of mean residual life (MRL) that leverage all available patient information, including longitudinal biomarker data. Biomarkers are biological, clinical, and other variables reflecting disease progression that are often measured repeatedly on patients in the clinical setting. Dynamic prediction methods leverage accruing biomarker measurements to improve performance, providing updated predictions as new measurements become available. We introduce two methods for dynamic prediction of MRL using longitudinal biomarkers. in both methods, we begin by using long short-term memory networks (LSTMs) to construct encoded representations of the biomarker trajectories, referred to as "context vectors." In our first method, the LSTM-GLM, we dynamically predict MRL via a transformed MRL model that includes the context vectors as covariates. In our second method, the LSTM-NN, we dynamically predict MRL from the context vectors using a feed-forward neural network. We demonstrate the improved performance of both proposed methods relative to competing methods in simulation studies. We apply the proposed methods to dynamically predict the restricted mean residual life (RMRL) of septic patients in the intensive care unit using electronic medical record data. We demonstrate that the LSTM-GLM and the LSTM-NN are useful tools for producing individualized, real-time predictions of RMRL that can help inform the treatment decisions of septic patients.

Dynamic and static predictive modelling of psychotherapy outcome-Comparison of two statistical approaches.

OBJECTIVE: Improving prediction abilities in the therapy process can increase therapeutic success for a variety of reasons, such as more personalised treatment or resource optimisation. The increasingly applied methods of dynamic prediction seem to be very promising for this purpose. Prediction models are usually based on static approaches of frequentist statistics. However, the application of this statistical approach has been widely criticised in this research area. Bayesian statistics has been proposed in the literature as an alternative, especially for the task of dynamic modelling. In this study, we compare the performance of predicting therapy outcome over the course of therapy between both statistical approaches. METHOD: Based on a sample of 341 patients, a logistic regression analysis was performed using both statistical approaches. Therapy success was conceptualised as reliable pre-post improvement in brief symptom inventory (BSI) scores. As predictors, we used the subscales of the Outcome Questionnaire (OQ-30) and the Helping Alliance Questionnaire (HAQ) measured every fifth session, as well as baseline BSI scores. RESULTS: The influence of the predictors during therapy differs between the frequentist and the Bayesian approach. In contrast, predictive validity is comparable with a mean area under the curve (AUC) of 0.76 in both model types. CONCLUSION: Bayesian statistic provides an innovative and useful alternative to the frequentist approach in predicting therapy outcome. The theoretical foundation is particularly well suited for dynamic prediction. Nevertheless, no differences in predictive validity were found in this study. More complex methodology as well as further research seems necessary to exploit the potential of Bayesian statistics in this area.

Random survival forests with multivariate longitudinal endogenous covariates.

Predicting the individual risk of clinical events using the complete patient history is a major challenge in personalized medicine. Analytical methods have to account for a possibly large number of time-dependent predictors, which are often characterized by irregular and error-prone measurements, and are truncated early by the event. In this work, we extended the competing-risk random survival forests to handle such endogenous longitudinal predictors when predicting event probabilities. The method, implemented in the R package DynForest, internally transforms the time-dependent predictors at each node of each tree into time-fixed features (using mixed models) that can then be used as splitting candidates. The final individual event probability is computed as the average of leaf-specific Aalen-Johansen estimators over the trees. Using simulations, we compared the performances of DynForest to accurately predict an event with (i) a joint modeling alternative when considering two longitudinal predictors only, and with (ii) a regression calibration method that ignores the informative truncation by the event when dealing with a large number of longitudinal predictors. Through an application in dementia research, we also illustrated how DynForest can be used to develop a dynamic prediction tool for dementia from multimodal repeated markers, and quantify the importance of each marker.