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.

Sensitivity Analysis for Survival Prognostic Prediction with Gene Selection: A Copula Method for Dependent Censoring

Prognostic analysis for patient survival often employs gene expressions obtained from high-throughput screening for tumor tissues from patients. When dealing with survival data, a dependent censoring phenomenon arises, and thus the traditional Cox model may not correctly identify the effect of each gene. A copula-based gene selection model can effectively adjust for dependent censoring, yielding a multi-gene predictor for survival prognosis. However, methods to assess the impact of various types of dependent censoring on the multi-gene predictor have not been developed. In this article, we propose a sensitivity analysis method using the copula-graphic estimator under dependent censoring, and implement relevant methods in the R package “compound.Cox”. The purpose of the proposed method is to investigate the sensitivity of the multi-gene predictor to a variety of dependent censoring mechanisms. In order to make the proposed sensitivity analysis practical, we develop a web application. We apply the proposed method and the web application to a lung cancer dataset. We provide a template file so that developers can modify the template to establish their own web applications.

Renal function decline and heart failure hospitalisation in patients with type 2 diabetes: Dynamic predictions from the prospective SURDIAGENE cohort

AIMS

For type 2 diabetes persons, we assessed the association between renal function decline and heart failure hospitalisation (HFH) and validated dynamic HFH predictions (DynHFH) based on repeated estimated Glomerular Filtration Rate (eGFR) values.

METHODS

We studied 1413 patients from the SURDIAGENE cohort. From a joint model for longitudinal CKD-EPI measures and HFH risk, we calculated the probability of being HFH-free in the next five years.

RESULTS

The mean eGFR decline was estimated at 1.48 ml/min/1.73 m² per year (95 % CI from 1.23 to 1.74). We observed that eGFR decline was significantly associated with the HFH risk (_adjHR = 1.15 for an increase in yearly decline of 1 ml/min/1.73 m², 95 % CI from 1.03 to 1.26) independently of 7 baseline variables (from clinical, biological and ECG domains). Discrimination was good along the prediction times: AUC at 0.87 (95 %CI from 0.84 to 0.91) at patient inclusion and 0.77 (95 %CI from 0.67 to 0.87) at seven years' follow-up.

CONCLUSIONS

Renal function decline was significantly associated with the HFH risk. In the era of computer-assisted medical decisions, the DynHFH, a tool that dynamically predicts HFH in type 2 diabetes persons (https://shiny.idbc.fr/DynHFH), might be helpful for precision medicine and the implementation of stratified medical decision-making.

Case Report: Long-term observations from the tacrolimus weaning randomized clinical trial depicts the challenging aspects for determination of low-immunological risk patients

Whilst calcineurin inhibitors (CNI) are the cornerstone of immunosuppressive maintenance therapy in kidney transplantation, several studies have investigated the safety of CNI withdrawal in order to avoid their numerous side effects. In this context, we performed several years ago a clinical randomized trial evaluating CNI weaning in stable kidney transplant recipients without anti-HLA immunization. The trial was interrupted prematurely due to a high number of de novo DSA (dnDSA) and biopsy proven acute rejection (BPAR) in patients who underwent tacrolimus weaning, resulting in treatment for rejection and resumption of tacrolimus. We report here the long-term outcomes of patients included in this clinical trial. Ten years after randomization, all patients are alive with a functional allograft. They all receive tacrolimus therapy except one with recurrent cutaneous neoplasia issues. Long-term eGFR was comparable between patients of the two randomized groups (46.4 ml/min vs 42.8 ml/min). All dnDSA that occurred during the study period became non-detectable and all rejections episodes were reversed. The retrospective assessment of HLA DQ single molecule epitope mismatching determined that a majority of patients who developed dnDSA after tacrolimus withdrawal would have been considered at high immunological risk. Minimization of immunosuppression remains a challenging objective, mainly because of the issues to properly select very low immunological risk patients. Valuable improvements have been made the last decade regarding evaluation of the allograft rejection notably through the determination of numerous at-risk biomarkers. However, even if the impact of such tools still need to be clarify in clinical routine, they may permit an improvement in patients' selection for immunosuppression minimization without increasing the risk of allograft rejection.

Individual dynamic prediction and prognostic analysis for long-term allograft survival after kidney transplantation

Background

Predicting allograft survival is vital for efficient transplant success. With dynamic changes in patient conditions, clinical indicators may change longitudinally, and doctors’ judgments may be highly variable. It is necessary to establish a dynamic model to precisely predict the individual risk/survival of new allografts.

Methods

The follow-up data of 407 patients were obtained from a renal allograft failure study. We introduced a landmarking-based dynamic Cox model that incorporated baseline values (age at transplantation, sex, weight) and longitudinal changes (glomerular filtration rate, proteinuria, hematocrit). Model performance was evaluated using Harrell’s C-index and the Brier score.

Results

Six predictors were included in our analysis. The Kaplan–Meier estimates of survival at baseline showed an overall 5-year survival rate of 87.2%. The dynamic Cox model showed the individual survival prediction with more accuracy at different time points (for the 5-year survival prediction, the C-index = 0.789 and Brier score = 0.065 for the average of all time points) than the static Cox model at baseline (C-index = 0.558, Brier score = 0.095). Longitudinal covariate prognostic analysis (with time-varying effects) was performed.

Conclusions

The dynamic Cox model can utilize clinical follow-up data, including longitudinal patient information. Dynamic prediction and prognostic analysis can be used to provide evidence and a reference to better guide clinical decision-making for applying early treatment to patients at high risk.

Predicting the Kidney Graft Survival Using Optimized African Buffalo-Based Artificial Neural Network

A variety of receptor and donor characteristics influence long-and short-term kidney graft survival. It is critical to predict the effectiveness of kidney transplantation to optimise organ allocation. This would allow patients to choose the best accessible kidney donor and the optimal immunosuppressive medication. Several studies have attempted to identify factors that predispose to graft rejection, but the results have been contradictory. As a result, the goal of this paper is to use the African buffalo-based artificial neural network (AB-ANN) approach to uncover predictive risk variables related to kidney graft. These two feature selection approaches combine to provide a novel hybrid feature selection technique that could select the most important elements to improve prediction accuracy. The feature analysis revealed that clinical features have varied effects on transplant survival. The collected data is processed in both training and testing methods. The prediction model's performance, in terms of accuracy, precision, recall, and F-measure, was examined, and the results were compared with those of other existing systems, including naive Bayesian, random forest, and J48 classifier. The results suggest that the proposed approach can forecast graft survival in kidney recipients' next visits in a creative manner and with more accuracy compared with other classifiers. This proposed method is more efficient for predicting kidney graft survival. Incorporating those clinical tools into outpatient clinics' everyday workflows could help physicians make better and more personalised decisions.

Dynamic Risk Prediction via a Joint Frailty-Copula Model and IPD Meta-Analysis: Building Web Applications

Clinical risk prediction formulas for cancer patients can be improved by dynamically updating the formulas by intermediate events, such as tumor progression. The increased accessibility of individual patient data (IPD) from multiple studies has motivated the development of dynamic prediction formulas accounting for between-study heterogeneity. A joint frailty-copula model for overall survival and time to tumor progression has the potential to develop a dynamic prediction formula of death from heterogenous studies. However, the process of developing, validating, and publishing the prediction formula is complex, which has not been sufficiently described in the literature. In this article, we provide a tutorial in order to build a web-based application for dynamic risk prediction for cancer patients on the basis of the R packages joint.Cox and Shiny. We demonstrate the proposed methods using a dataset of breast cancer patients from multiple clinical studies. Following this tutorial, we demonstrate how one can publish web applications available online, which can be manipulated by any user through a smartphone or personal computer. After learning this tutorial, developers acquire the ability to build an online web application using their own datasets.

Dynamic prediction models for graft failure in paediatric kidney transplantation

BACKGROUND

Several models have been proposed to predict kidney graft failure in adult recipients but none in younger recipients. Our objective was to propose a dynamic prediction model for graft failure in young kidney transplant recipients.

METHODS

We included 793 kidney transplant recipients waitlisted before the age of 18 years who received a first kidney transplantation before the age of 21 years in France in 2002-13 and survived >90 days with a functioning graft. We used a Cox model including baseline predictors only (sex, age at transplant, primary kidney disease, dialysis duration, donor type and age, human leucocyte antigen matching, cytomegalovirus serostatus, cold ischaemia time and delayed graft function) and two joint models also accounting for post-transplant estimated glomerular filtration rate (eGFR) trajectory. Predictive performances were evaluated using a cross-validated area under the curve (AUC) and R2 curves.

RESULTS

When predicting the risk of graft failure from any time within the first 7 years after paediatric kidney transplantation, the predictions for the following 3 or 5 years were accurate and much better with the joint models than with the Cox model (AUC ranged from 0.83 to 0.91 for the joint models versus 0.56 to 0.64 for the Cox model).

CONCLUSION

Accounting for post-transplant eGFR trajectory strongly increased the accuracy of graft failure prediction in young kidney transplant recipients.

External Validation of the DynPG for Kidney Transplant Recipients

BACKGROUND

In kidney transplantation, dynamic prediction of patient and kidney graft survival (DynPG) may help to promote therapeutic alliance by delivering personalized evidence-based information about long-term graft survival for kidney transplant recipients. The objective of the current study is to externally validate the DynPG.

METHODS

Based on 6 baseline variables, the DynPG can be updated with any new serum creatinine measure available during the follow-up. From an external validation sample of 1637 kidney recipients with a functioning graft at 1-year posttransplantation from 2 European transplantation centers, we assessed the prognostic performance of the DynPG.

RESULTS

As one can expect from an external validation sample, differences in several recipient, donor, and transplantation characteristics compared with the learning sample were observed. Patients were mainly transplanted from deceased donors (91.6% versus 84.8%; P < 0.01), were less immunized against HLA class I (18.4% versus 32.7%; P < 0.01) and presented less comorbidities (62.2% for hypertension versus 82.7%, P < 0.01; 25.1% for cardiovascular disease versus 33.9%, P < 0.01). Despite these noteworthy differences, the area under the ROC curve varied from 0.70 (95% confidence interval [CI], 0.64-0.76) to 0.76 (95% CI, 0.64-0.88) for prediction times at 1 and 6 years posttransplantation respectively, and calibration plots revealed reasonably accurate predictions.

CONCLUSIONS

We validated the prognostic capacities of the DynPG in terms of both discrimination and calibration. Our study showed the robustness of the DynPG for informing both the patient and the physician, and its transportability for a cohort presenting different features than the one used for the DynPG development.

Prediction scores for risk of allograft loss in patients receiving kidney transplants: nil satis nisi optimum

Long-term graft survival is the main concern of kidney transplantation. Some strategies have been tested to predict graft survival using estimated glomerular filtration rate or proteinuria at different time points, histologic assessment, non-invasive biomarkers or even machine-learning methods. However, the 'magical formulae' for allograft survival prediction does not exist yet.

Prognostic value for long-term graft survival of estimated glomerular filtration rate and proteinuria quantified at 3 months after kidney transplantation

Background

The estimated glomerular filtration rate (eGFR) measured at 1 year is the usual benchmark applied in kidney transplantation (KT). However, acting on earlier eGFR values could help in managing KT during the first post-operative year. We aimed to assess the prognostic value for long-term graft survival of the early (3 months) quantification of eGFR and proteinuria following KT.

Methods

The 3-, 6- and 12-month eGFR using the Modified Diet in Renal Disease equation (eGFR_MDRD) was determined and proteinuria was measured in 754 patients who underwent their first KT between 2000 and 2010 (with a mean follow-up of 8.3 years) in our centre. Adjusted associations with graft survival were estimated using a multivariable Cox model. The predictive accuracy was estimated using the C-index and net reclassification index. These same analyses were measured in a multicentre validation cohort of 1936 patients.

Results

Both 3-month eGFR_MDRD and proteinuria were independent predictors of return to dialysis (all P < 0.05) and there was a strong correlation between eGFR at 3 and 12 months (Spearman’s ρ = 0.76). The predictive accuracy of the 3-month eGFR was within a similar range and did not differ significantly from the 12-month eGFR in either the derivation cohort [C-index 62.6 (range 57.2–68.1) versus 66.0 (range 60.1–71.9), P = 0.41] or the validation cohort [C-index 69.3 (range 66.4–72.1) versus 71.7 (range 68.7–74.6), P = 0.25].

Conclusion

The 3-month eGFR was a valuable predictor of the long-term return to dialysis whose predictive accuracy was not significantly less than that of the 12-month eGFR in multicentre cohorts totalling >2500 patients. Three-month outcomes may be useful in randomized controlled trials targeting early therapeutic interventions.