Clinical Decision Support

When a novel marker is developed, it is essential to understand how it can be integrated with existing clinicopathological parameters commonly used for treatment-related decisions, if it may be strengthened by combining with other markers and how it can be made available to the clinician. We have established a collaboration with DIPS, the most common electronic patient record (EPR) system in Norway, for implementation of our solutions for decision support.

The clinical decision support project is about identifying an optimal combination of prognostic parameters and implementing them into an electronic patient record system. With several new and existing prognostic markers, a system for clinical decision support with a simple output is essential for the clinicians to be able to use our new methods. Data from our projects on prostate- and colorectal cancer will be used to develop the methodology. We have established a collaboration with DIPS for implementation of our solutions.

Integration with the EPR system requires interaction with several modules in the existing infrastructure to extract, e.g., image data and link our analysis results with existing parameters. Furthermore, a range of new modules is required, including an image viewer for pathology images and a module to request analyses, as well as modules for performing the actual analyses). Some analyses will be available as commercial services, and some analyses are only valid for specific cancer types or patient subgroups; these properties require a design where the available analyses are relative to the patient profile.

Training dataset

The first training dataset is ready, based on the QUASAR 2 cohort. Of 1941 patients analysed in the original trial report, it appears that 1379 consented to the use of tissue samples and did not withdraw their follow-up consent. A death was recorded for 262 (19%) of the 1379 patients, of whom 187 died from CRC. The median follow-up of patients alive at last follow-up was 5.0 years (interquartile range 4.0 to 5.2 years). A recurrence of the CRC was recorded for 322 (23%) of the 1379 patients. Data are integrated or compared from about 20 source files into a large database. This database contains data on all candidate markers that are currently available.

Project plans

1. Finish the data consolidation process by combining some variables in a similar manner as done previously, making a short description, and reducing the data to one observation per patient.
2. Randomly divide into a training subset and a test subset - use 75% for training and 25% for testing and perform the random sampling when stratified on pN stage. An alternative would be to not do any stratification but instead compare the distribution of many candidate markers in the two subsets.
3. Use multivariable Cox regression analysis to design a clinical decision support system, possibly one for stage II and one for stage III colorectal cancer. We plan for the system to provide a score that relates to the clinical outcome, particularly the probability of cancer-specific death. It will likely be a continuous score, although it is possible to discretise it, which was done, e.g., in the development of the Cancer of the Prostate Risk Assessment Postsurgical (CAPRA-S) score.
4. Enable prediction in the absence of certain or any markers. This can be done by developing separate clinical decision support systems, which might be interesting to evaluate the general importance of individual markers. It may also be done at the time of inference, which might be necessary to avoid designing and evaluating a multitude of systems and will also provide estimates that are more stable when sequentially including additional markers in a clinical application of the system. Appropriate approaches for doing this during inference needs further consideration. One possibility is to assay the full system for every possible value of the missing markers and then combine the estimated probability distributions for a particular clinical outcome (e.g., cancer-specific death), possibly using weighted averaging where the weights are proportional to an estimated probability of the missing markers being those values (thus in a sense resembling imputation strategies).
5. Validate a selected clinical decision support system, or one for each stage, on external cohort(s).