| Summary: | 碩士 === 國立臺灣大學 === 流行病學與預防醫學研究所 === 104 === Background
While receiver operating characteristics (ROC) curve has been widely used to evaluate the performance of diagnostic tool its application to evaluate the performance of screening tool used in population-based cancer screening is not straightforward because the identification of asymptomatic cancer, staying in pre-clinical screen-detectable phase (PCDP), is often faced with incomplete ascertainment that is highly dependent on follow-up time and mean sojourn time (MST) of cancer in question, representing the rate of disease progression from PCDP to CP (clinical phase).
We are motivated by evaluating the performance of one community-based and the expanded nationwide population-based screening program in Taiwan. Two issues indicated above are encountered. To solve these two, different studies or single study with different follow-up times can be handled in the context of meta-analysis of ROC. The ROC curve adjusting for MST is solved by the combination of multistate stochastic process with the developed meta-analytic ROC method.
Aims
1. We used summary ROC (SROC) and hierarchical ROC (HROC) curve with different estimation methods, including Monte Carlo Markov Chan (MCMC), generalized linear model with maximum likelihood estimation (MLE), and non-linear mixed model with moment method (ME) to evaluate the performance of population-based colorectal cancer screening to deal with incomplete ascertainment due to different follow-up times;
2. We then developed Bayesian directed graphic model to link both diagnostic accuracy and threshold encoded in the context of SROC underpinning at study level with the random outcome of true positive and false positive at individual level;
3. We then combined Bayesian DAG model under the context of SROC with multistate Markov underpinning to derived mean sojourn time(MST)-adjusted ROC.
Material and Methods
Three sources of dataset were used for the analysis of ROC for FIT in the CRC screening. First, primary data on FIT screening embedded in the Keelung Community-based Integrated Screening (KCIS) were used. The second data are also primary data but derived from the nationwide FIT screening in Taiwan. The third source for meta-analysis was derived from published data, of which faecal-based CRC screening, including both guaic-feacal occult blood test and FIT, and scopy-based CRC screening programs were enrolled. For the two primary dataset, we kept the quantitative value of FIT for all subjects, including disease-free subjects, screen-detected and interval cancers. Therefore, a series of sensitivity and specificity with varying cut-off for positive FIT were derived for the further ROC analysis.
We firstly depicted the crude ROC curve in the KCIS data with FIT cut-off levels at 20, 40, 60, 80, 100, 150, 250, 450 ng/ml, whereas interval cancers occurring within 24 months after previous negative screen was treated as false negative. The Youden index was used to assess the optimal cut-off level. The parametric methods of summary ROC (SROC) and hierarchical SROC (HSROC) were applied to assess the behavior of SROC. Summary measures of diagnostic odds ratio, threshold, and AUC given homogeneity between studies were obtained.
A Bayesian directed acyclic graphic (DAG) model for SROC was developed to tackle the problems of SROC and HSROC, in which the diagnostic log-odds ratio and diagnostic threshold were deterministic, and the difficulty of obtain AUC when the assumption of heterogeneous studies was violated.
A further model which takes into account the possible false negative cases for those with FIT level lower than the cut-off levels by combing a multi-state model for the disease natural history for CRC was developed under the context of Bayesian DAG model. With the corrected number of false negative cases, we can obtain the multi-state summary ROC (MSROC) .
All the methods were also applied to the Taiwanese nationwide CRC screening data. A final meta-analysis incorporating selected studies from literature gives the pooled results of ROC for FIT screening given the current state-of-art evidence.
Main Results and Conclusions
From the practical aspect of population-based colorectal cancer screening with FIT, evaluating performance of FIT test with the developed methodologies mentioned above can throw light on several interesting findings.
1. The performance of FIT in population-based CRC screening evaluated with empirical ROC curve yields too optimistic results, 87% AUChom and 80% Q* without considering follow-up time and adjusting for mean sojourn time.
2. The performance of FIT in population-based CRC screening evaluated with summary and hierarchical summary ROC curve varies with follow-up time and MST.
3. Assuming all interval cancers arising from false negative cases give too conservative results, 71% AUChom and 66% Q*.
4. The corrected ROC curve with refined parameters from three-state Markov process considering interval cancer composed of false negative cases and newly incident cases gives base-case estimates, 79%AUChom and 73% Q*.
From the viewpoint of methodology
This thesis first proposed summary ROC and hierarchical SROC in combination with interval-cancer-based follow-up study design to evaluate performance of screening tool used in population-based cancer screening with the consideration of time dimension of follow-up time. We then develop Bayesian directed acyclic graphic (DAG) model to estimate the parameters defined under the framework of SROC and HSROC with great flexibility of considering observed and unobserved heterogeneity corresponding to covariates (fixed effects) and random-effect (the incorporation of random-effect. The third step was to develop a novel dynamic multistate SROC and HSROC method with Bayesian DAG underpinning to correct both curves and relevant parameters with the estimated preclinical incidence rate and mean sojourn time (MST) obtained from a three-state Markov process. Applications of the proposed Bayesian DAG MSROC model to population-based colorectal cancer screening can be helpful for calibrating SROC and HSROC curves for unbiased evaluation of performance of FIT test in population-based screening
|
|---|
