Prospective Validation of an EHR-based Pancreatic Cancer Risk Model
Study Details
Study Description
Brief Summary
The goal of this prospective observational cohort study is to validate a previously developed pancreatic cancer risk prediction algorith (the PRISM model) using electronic health records from the general population. The main questions it aims to answer are:
-
Will a pancreatic cancer risk model, developed on routine EHR data, reliably and accurately predict pancreatic cancer in real-time?
-
What is the average time from model deployment and risk prediction, to the date of pancreatic cancer development and what is the stage of pancreatic cancer at diagnosis? The risk model will be deployed on data from individuals eligible for the study. Each individual will be assigned a risk score and tracked over time to assess the model's discriminatory performance and calibration.
| Condition or Disease | Intervention/Treatment | Phase |
|---|---|---|
|
Detailed Description
To prospectively validate, implement in real-time, and assess performance of an EHR- based PDAC risk-prediction model. To test the hypothesis that our model will reliably predict PDAC in a real-time clinical setting, we will conduct a multi-center prospective cohort study, deploying the PDAC risk model within the TriNetX federated network database, and will take the following steps:
- generate a risk prediction score for each individual under the care of 44 health care organizations (HCOs) in the USA ii) follow all individuals for up to 3 years to assess the primary end-point of PDAC development.
The following metrics will be used to test the discriminative performance and calibration of the EHR-based PDAC risk model in predicting incident PDAC, at the end of the 3-year period:
-
AUROC, sensitivity, specificity, PPV/NPV for assessing discrimination
-
Calibration: for assessing the accuracy of estimates, based on the estimated to observed number of events
Study Design
Arms and Interventions
| Arm | Intervention/Treatment |
|---|---|
| prospective general opulation cohort Males and females age >= 40 years, without a personal history of PDAC or current PDAC, with at least 2 clinical encounters to the HCO within the year prior to the study start date. |
Other: Pancreatic Cancer Risk Model (PRISM)
A neural network model (PrismNN) and a logistic regression model (PrismLR) that use routinely collected EHR data to stratify individuals from the general population into PDAC risk groups
|
Outcome Measures
Primary Outcome Measures
- Area under the receiver operating characteristic curve (AUROC) of PRISM for all groups stratified [6 months from index date]
To assess the discriminatory performance of PRISM for prospective identification of high-risk individuals for PDAC development. ROCs and AUROC numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location.
- Area under the receiver operating characteristic curve (AUROC) of PRISM for all groups stratified [at 1 year]
To assess the discriminatory performance of PRISM for prospective identification of high-risk individuals for PDAC development. ROCs and AUROC numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location.
- Area under the receiver operating characteristic curve (AUROC) of PRISM for all groups stratified [at 2 years]
To assess the discriminatory performance of PRISM for prospective identification of high-risk individuals for PDAC development. ROCs and AUROC numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location.
- Area under the receiver operating characteristic curve (AUROC) of PRISM for all groups stratified [at 3 years]
To assess the discriminatory performance of PRISM for prospective identification of high-risk individuals for PDAC development. ROCs and AUROC numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location.
- Calibration of PRISM for all groups stratified [6 months from index date]
To access how well the risk prediction by PRISM aligns with observed risk without recalibration. Calibration plots will be created for the whole population and groups stratified by age, sex, race, and geographical location.
- Calibration of PRISM for all groups stratified [at 1 year]
To access how well the risk prediction by PRISM aligns with observed risk without recalibration. Calibration plots will be created for the whole population and groups stratified by age, sex, race, and geographical location.
- Calibration of PRISM for all groups stratified [at 2 years]
To access how well the risk prediction by PRISM aligns with observed risk without recalibration. Calibration plots will be created for the whole population and groups stratified by age, sex, race, and geographical location.
- Calibration of PRISM for all groups stratified [at 3 years]
To access how well the risk prediction by PRISM aligns with observed risk without recalibration. Calibration plots will be created for the whole population and groups stratified by age, sex, race, and geographical location.
- PRISM performance metrics of high-risk group for direct screening [6 months from index date]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as high-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined based on Standardized Incidence Ratio of 5 or greater. The absolute one-year risk thresholds are 0.1834% for PrismNN and 0.2048% risk for PrismLR. SIR 5 or greater was chosen because it is comparable to the current screening inclusion threshold for individuals with an inherited predisposition.
- PRISM performance metrics of high-risk group for direct screening [at 1 year]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as high-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined based on Standardized Incidence Ratio of 5 or greater. The absolute one-year risk thresholds are 0.1834% for PrismNN and 0.2048% risk for PrismLR. SIR 5 or greater was chosen because it is comparable to the current screening inclusion threshold for individuals with an inherited predisposition.
- PRISM performance metrics of high-risk group for direct screening [at 2 years]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as high-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined based on Standardized Incidence Ratio of 5 or greater. The absolute one-year risk thresholds are 0.1834% for PrismNN and 0.2048% risk for PrismLR. SIR 5 or greater was chosen because it is comparable to the current screening inclusion threshold for individuals with an inherited predisposition.
- PRISM performance metrics of high-risk group for direct screening [at 3 years]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as high-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined based on Standardized Incidence Ratio of 5 or greater. The absolute one-year risk thresholds are 0.1834% for PrismNN and 0.2048% risk for PrismLR. SIR 5 or greater was chosen because it is comparable to the current screening inclusion threshold for individuals with an inherited predisposition.
- PRISM performance metrics of medium-risk group for biomarker testing [6 months from index date]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as medium-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined by specificity 85%, with sensitivity around 46%. The absolute one-year risk thresholds are 0.0574% for PrismNN and 0.0564% for PrismLR. Prism operates as a tiered system for identifying individuals in need of screening with this lower risk threshold.
- PRISM performance metrics of medium-risk group for biomarker testing [at 1 year]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as medium-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined by specificity 85%, with sensitivity around 46%. The absolute one-year risk thresholds are 0.0574% for PrismNN and 0.0564% for PrismLR. Prism operates as a tiered system for identifying individuals in need of screening with this lower risk threshold.
- PRISM performance metrics of medium-risk group for biomarker testing [at 2 years]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as medium-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined by specificity 85%, with sensitivity around 46%. The absolute one-year risk thresholds are 0.0574% for PrismNN and 0.0564% for PrismLR. Prism operates as a tiered system for identifying individuals in need of screening with this lower risk threshold.
- PRISM performance metrics of medium-risk group for biomarker testing [at 3 years]
To evaluate sensitivity, specificity, PPV, and SIR for patients identified as medium-risk by PRISM. Numbers will be calculated for the whole population and groups stratified by age, sex, race, and geographical location. Risk threshold determined by specificity 85%, with sensitivity around 46%. The absolute one-year risk thresholds are 0.0574% for PrismNN and 0.0564% for PrismLR. Prism operates as a tiered system for identifying individuals in need of screening with this lower risk threshold.
Secondary Outcome Measures
- Timing of incident PDAC occurrence [6 months from index date]
To evaluate how long in advance before PDAC occurrence should be expected for PRISM models to make high-risk predictions. Distribution plots of the date of PDAC incidence for high/medium-risk groups (defined in Outcome 3 and Outcome 4) will be created.
- Timing of incident PDAC occurrence [at 1 year]
To evaluate how long in advance before PDAC occurrence should be expected for PRISM models to make high-risk predictions. Distribution plots of the date of PDAC incidence for high/medium-risk groups (defined in Outcome 3 and Outcome 4) will be created.
- Timing of incident PDAC occurrence [at 2 years]
To evaluate how long in advance before PDAC occurrence should be expected for PRISM models to make high-risk predictions. Distribution plots of the date of PDAC incidence for high/medium-risk groups (defined in Outcome 3 and Outcome 4) will be created.
- Timing of incident PDAC occurrence [at 3 years]
To evaluate how long in advance before PDAC occurrence should be expected for PRISM models to make high-risk predictions. Distribution plots of the date of PDAC incidence for high/medium-risk groups (defined in Outcome 3 and Outcome 4) will be created.
- Tumor stage at PDAC diagnosis [6 months from index date]
stage at diagnosis per tumor registry/pathology report
- Tumor stage at PDAC diagnosis [at 1 year]
stage at diagnosis per tumor registry/pathology report
- Tumor stage at PDAC diagnosis [at 2 years]
stage at diagnosis per tumor registry/pathology report
- Tumor stage at PDAC diagnosis [at 3 years]
stage at diagnosis per tumor registry/pathology report
Eligibility Criteria
Criteria
Inclusion Criteria:
-
Male and females age >= 40 years from 44 US HCOs from the TriNetX platform
-
at least 2 clinical encounters to the HCO, within the last year, before the study start date
Exclusion Criteria:
-
Personal history of PDAC or current PDAC
-
Age below 40
Notes on sampling method: no sampling was performed. All eligible individuals are included in this study.
Contacts and Locations
Locations
| Site | City | State | Country | Postal Code | |
|---|---|---|---|---|---|
| 1 | Beth Israel Deaconess Medical Center | Boston | Massachusetts | United States | 02115 |
Sponsors and Collaborators
- Beth Israel Deaconess Medical Center
- Massachusetts Institute of Technology
- TriNetX, LLC
Investigators
- Principal Investigator: Limor Appelbaum, MD, Beth Israel Deaconess Medical Center
Study Documents (Full-Text)
None provided.More Information
Publications
None provided.- 2023Trial