Search

Currently available risk prediction methods are limited in their ability to deal with complex, heterogeneous, and longitudinal data such as that available in primary care records, or in their ability to deal with multiple competing risks. This paper develops a novel deep learning approach that is able to successfully address current limitations of standard statistical approaches such as landmarking and joint modeling. Our approach, which we call Dynamic-DeepHit, flexibly incorporates the available longitudinal data comprising various repeated measurements (rather than only the last available measurements) in order to issue dynamically updated survival predictions for one or multiple competing risk(s). Dynamic-DeepHit learns the time-to-event distributions without the need to make any assumptions about the underlying stochastic models for the longitudinal and the time-to-event processes. Thus, unlike existing works in statistics, our method is able to learn data-driven associations between the longitudinal data and the various associated risks without underlying model specifications. We demonstrate the power of our approach by applying it to a real-world longitudinal dataset from the U.K. Cystic Fibrosis Registry, which includes a heterogeneous cohort of 5883 adult patients with annual follow-ups between 2009 to 2015. The results show that Dynamic-DeepHit provides a drastic improvement in discriminating individual risks of different forms of failures due to cystic fibrosis. Furthermore, our analysis utilizes post-processing statistics that provide clinical insight by measuring the influence of each covariate on risk predictions and the temporal importance of longitudinal measurements, thereby enabling us to identify covariates that are influential for different competing risks. [ABSTRACT FROM AUTHOR]

Background: Recommended cardiovascular disease (CVD) prediction tools do not account for competing mortality risk and over-predict incident CVD in older and multimorbid people. The aim of this study was to derive and validate a competing risk model (CRISK) to predict incident CVD and compare its performance to that of QRISK3 in UK primary care.Methods: We used UK linked primary care data from the Clinical Practice Research Datalink (CPRD) GOLD to identify people aged 25-84 years with no previous CVD or statin treatment split into derivation and validation cohorts. In the derivation cohort, we derived models using the same covariates as QRISK3 with Fine-Gray competing risk modelling alone (CRISK) and with Charlson Comorbidity score (CRISK-CCI) as an additional predictor of non-CVD death. In a separate validation cohort, we examined discrimination and calibration compared to QRISK3. Reclassification analysis examined the number of patients recommended for treatment and the estimated number needed to treat (NNT) to prevent a new CVD event.Results: The derivation and validation cohorts included 989,732 and 494,865 women and 946,784 and 473,392 men respectively. Overall discrimination of CRISK and CRISK-CCI were excellent and similar to QRISK3 (for women, C-statistic = 0.863/0.864/0.863 respectively; for men 0.833/0.819/0.832 respectively). CRISK and CRISK-CCI calibration overall and in younger people was excellent. CRISK over-predicted in older and multimorbid people although performed better than QRISK3, whilst CRISK-CCI performed the best. The proportion of people reclassified by CRISK-CCI varied by QRISK3 risk score category, with 0.7-9.7% of women and 2.8-25.2% of men reclassified as higher risk and 21.0-69.1% of women and 27.1-57.4% of men reclassified as lower risk. Overall, CRISK-CCI recommended fewer people for treatment and had a lower estimated NNT at 10% risk threshold. Patients reclassified as higher risk were younger, had lower SBP and higher BMI, and were more likely to smoke.Conclusions: CRISK and CRISK-CCI performed better than QRISK3. CRISK-CCI recommends fewer people for treatment and has a lower NNT to prevent a new CVD event compared to QRISK3. Competing risk models should be recommended for CVD primary prevention treatment recommendations. [ABSTRACT FROM AUTHOR]

Background: The increasing incidence of diffuse large B-cell lymphoma (DLBCL) in ageing populations places a significant burden on healthcare systems. Co-morbidity, frailty, and reduced organ and physiological reserve contribute to treatment-related complications. The optimal dose intensity of R-CHOP to optimize outcome across different ages with variable frailty and comorbidity burden is unclear.Objectives and Methods: We examined the influence of intended (IDI) and relative (RDI) dose intensity of the combination of cyclophosphamide and doxorubicin, age and comorbidity on outcomes for DLBCL patients ≥70 years in a representative, consecutive cohort across eight UK centres (2009-2018). We determined predictors of survival using multivariable Cox regression, and predictors of recurrence before death using competing risks regression.Results: Porgression-free survival (PFS) and overall survival (OS) were significantly inferior in patients ≥80 vs. 70-79 years (P < 0.001). In contrast, 2-year cumulative relapse incidence, when accounting for non-relapse mortality as a competing risk, was no different between 70-79 vs. ≥80 years (P = 0.27) or comorbidity status (CIRS-G: 0-6 vs. >6) (P = 0.27). In 70-79 years, patients with an IDI ≥80% had a significantly improved PFS and OS (P < 0.001) compared to IDI < 80%. Conversely, in patients ≥80 years, there was no difference in PFS (P = 0.88) or OS (P = 0.75) according to IDI <80% vs. ≥80%. On multivariable analysis, when comparing by age, there was a significantly higher cumulative relapse rate for patients aged 70-79 years with an IDI <80% (vs. >80%) (P = 0.04) but not for patients ≥80 years comparing IDI (P = 0.32).Conclusion: 'R-mini-CHOP' provides adequate lymphoma-specific disease control and represents a reasonable treatment option in elderly patients ≥80 years aiming for cure. [ABSTRACT FROM AUTHOR]

Background: Continuous monitoring of surgical outcomes after joint replacement is needed to detect which brands' components have a higher than expected failure rate and are therefore no longer recommended to be used in surgical practice. We developed a monitoring method based on cumulative sum (CUSUM) chart specifically for this application.Methods: Our method entails the use of the competing risks model with the Weibull and the Gompertz hazard functions adjusted for observed covariates to approximate the baseline time-to-revision and time-to-death distributions, respectively. The correlated shared frailty terms for competing risks, corresponding to the operating unit, are also included in the model. A bootstrap-based boundary adjustment is then required for risk-adjusted CUSUM charts to guarantee a given probability of the false alarm rates. We propose a method to evaluate the CUSUM scores and the adjusted boundary for a survival model with the shared frailty terms. We also introduce a unit performance quality score based on the posterior frailty distribution. This method is illustrated using the 2003-2012 hip replacement data from the UK National Joint Registry (NJR).Results: We found that the best model included the shared frailty for revision but not for death. This means that the competing risks of revision and death are independent in NJR data. Our method was superior to the standard NJR methodology. For one of the two monitored components, it produced alarms four years before the increased failure rate came to the attention of the UK regulatory authorities. The hazard ratios of revision across the units varied from 0.38 to 2.28.Conclusions: An earlier detection of failure signal by our method in comparison to the standard method used by the NJR may be explained by proper risk-adjustment and the ability to accommodate time-dependent hazards. The continuous monitoring of hip replacement outcomes should include risk adjustment at both the individual and unit level. [ABSTRACT FROM AUTHOR]