Surveillance-dependent outcomes (SDOs) are continuous time failure outcomes that are identified by measurements that occur at discrete times such as patient visits or diagnostic exams. SDOs are ubiquitous in medical research. For example, many types of cancer progression are identified by surveillance screening. PSA recurrence in prostate cancer is identified by a rising PSA value on surveillance tests, and breast cancer recurrence after an initial diagnosis of in-situ disease is identified by surveillance mammography. In HIV research, the disease progression is identified by immune metrics such as viral load or CD4 counts collected at discrete times. And electronic health records provide patient visit records that can be used as sources of information on SDOs in observational studies.
SDOs are discrete snapshots of an underlying continuous time process, and may be subject to measurement error. As such, traditional analyses based on the observed SDOs rather than the underlying continuous-time process are limited in their generalizability and may be subject to bias. Cross-study comparisons are limited if the surveillance frequency differs between studies or if the studies are based on biomarkers with differential measurement error. Studying the underlying trajectory rather than the observed SDO process is more likely to provide insight into the disease process at the individual patient level and should facilitate meta-analysis across treatments or studies with different surveillance schema.
Our goal at the Fred Hutch SDO Working Group is to develop statistical tools to analyze SDOs and to apply them to real data in clinically relevant studies.
In a study of active surveillance in prostate cancer, we have used these tools to reconstruct the underlying progression trajectories, enabling comparisons across diverse observation schemes and accommodating measurement error in prostate biopsies. In the future we plan to expand our work to the study of cancer recurrence and metastasis after diagnosis. We are actively seeking collaborations across the center in this pervasive and important area of failure time research that is likely to have broad applicability in cancer studies.
R package cthmm: A package for fitting discretely observed continuous time hidden Markov models
Documentation: cthmm vignette
Lange JM, Minin VN. Fitting and Interpreting Continuous-Time Latent Markov Models for Panel Data. Statistics in medicine. 2013;32(26):4581-4595. doi:10.1002/sim.5861.
Lange, J. M., Hubbard, R. A., Inoue, L. Y. T. and Minin, V. N. (2015), A joint model for multistate disease processes and random informative observation times, with applications to electronic medical records data. Biom, 71: 90–101. doi:10.1111/biom.12252
Fred Hutchinson Cancer Research Center Biostatistics Seminar 3/22/2017