Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Machine learning risk stratification in patients with atherosclerotic cardiovascular disease: a personalized approach
Patients with atherosclerotic cardiovascular disease (ASCVD) are at high risk of suffering from future ASCVD events despite contemporary preventive therapies, and this risk has been termed “residual risk”. While the overall incidence of ASCVD has declined, the greatest decline has been in pre-hospital fatal events. Even with recent advances in cardiovascular interventions including statins, dual antiplatelet therapies, revascularization, and other therapies, ASCVD event rates among patients with prior ASCVD remain high (>3% per year). Furthermore, patients with ASCVD have a range of risks, yet personalized risk stratification tools are lacking.
Keywords: Machine learning for healthcare, risk prediction, large biobank, electronic health records data, cardiovascular risk prediction model
**Project goals:**
1. To develop the ML-based RRS score for predicting risk of recurrent ASCVD events. We will examine patients with prior ASCVD enrolled in the population-based UK Biobank prospective cohort (training subset sampled from N=33,949 patients with prior stroke or MI at enrollment; 10,007 recurrent ASCVD cases during mean follow-up of 10 years) to develop the ML-based RRS score to predict recurrent events. We will also examine the separate endpoints of predicting stroke (N=4,344 incident cases), myocardial infarction (N=2,149 cases), and ASCVD death (N=3,514). The RRS will be based on demographic variables, prior medical history, clinical risk factors, medication use, and ASCVD risk biomarkers. Data will be split into training and test sets to derive and internally validate the model.
2. To evaluate and validate the ML-based RRS score in patients with ASCVD from diverse populations. We will validate the RRS score using external and independent populations of patients with ASCVD: 1) validation subset of UK Biobank; 2) one more dataset, possibly a completed randomized clinical trial or other cohort study. We will assess model discrimination and calibration and compare RRS model performance with the clinical risk stratification algorithm of the 2018 cholesterol guidelines.
3.Compare the performance of the developed and validated ML model to the traditional risk prediction model developed in a separate project.
**Prerequisites:**
1. Experienced in Python and/or R programming
2. Familiarity or willingness to learn traditional risk prediction models and approaches: Cox proportional hazards model, competing risk analysis; evaluation of risk prediction model for clinical research: focus on calibration, net benefit, other performance statistics
3. Experience in working with machine learning models: elastic net, XGBoost, neural networks
4. Interest in learning about and adapting to the needs of clinical cardiovascular medical research
5. Basic familiarity with LINUX
**Work plan:**
1. Perform a literature review of risk prediction in clinical research: review existing classical models used for predicting the risk of primary cardiovascular events (ATPIII model, current PCE model, SCORE, AGLA and current secondary prevention schema), machine learning models applied to analysis of prospective cohorts and electronic medical records data. Epidemiological aspects of analysis: confounding, biases and imprecision specific to prospective cohorts, electronic medical records data, biobank data.
2. Some data preparation will be required, coordinated with data preparation and harmonization done at BWH, Boston, MA.
3. Develop a machine learning model of secondary ASCVD.
4. Compare its performance to classical risk prediction model developed in a separate project.
**Grading:**
- To receive a 6 the student must: meet the general goals and timeline of the thesis; work independently, demonstrate curiosity in the thesis topic and contribute her/his own ideas; communicate intermediate results clearly; write detailed and clear intermediate and final reports; document well project code; give a clear final presentation; work is exceptional and at a similar level expected for acceptance at leading international conferences or peer-reviewed journals.
- To receive a 5 the student must: meet the general goals of the thesis; work independently; write a detailed and clear final report, and give a clear final presentation; document well project code.
- To receive a 4 the student must: partially meet the general goals of the thesis; work somewhat independently; write a satisfactory final report, and give an understandable final presentation; document well project code
**Supervisor:**
Prof. Dr. Gunnar Rätsch D-INFK
**Advisors:**
Dr. Olga Demler, D-INFK/ Brigham and Women’s Hospital, Harvard Medical School
Dr. Samia Mora, Brigham and Women’s Hospital, Harvard Medical School
Olga Mineeva, D-INFK
**Project goals:** 1. To develop the ML-based RRS score for predicting risk of recurrent ASCVD events. We will examine patients with prior ASCVD enrolled in the population-based UK Biobank prospective cohort (training subset sampled from N=33,949 patients with prior stroke or MI at enrollment; 10,007 recurrent ASCVD cases during mean follow-up of 10 years) to develop the ML-based RRS score to predict recurrent events. We will also examine the separate endpoints of predicting stroke (N=4,344 incident cases), myocardial infarction (N=2,149 cases), and ASCVD death (N=3,514). The RRS will be based on demographic variables, prior medical history, clinical risk factors, medication use, and ASCVD risk biomarkers. Data will be split into training and test sets to derive and internally validate the model. 2. To evaluate and validate the ML-based RRS score in patients with ASCVD from diverse populations. We will validate the RRS score using external and independent populations of patients with ASCVD: 1) validation subset of UK Biobank; 2) one more dataset, possibly a completed randomized clinical trial or other cohort study. We will assess model discrimination and calibration and compare RRS model performance with the clinical risk stratification algorithm of the 2018 cholesterol guidelines. 3.Compare the performance of the developed and validated ML model to the traditional risk prediction model developed in a separate project.
**Prerequisites:** 1. Experienced in Python and/or R programming 2. Familiarity or willingness to learn traditional risk prediction models and approaches: Cox proportional hazards model, competing risk analysis; evaluation of risk prediction model for clinical research: focus on calibration, net benefit, other performance statistics 3. Experience in working with machine learning models: elastic net, XGBoost, neural networks 4. Interest in learning about and adapting to the needs of clinical cardiovascular medical research 5. Basic familiarity with LINUX
**Work plan:** 1. Perform a literature review of risk prediction in clinical research: review existing classical models used for predicting the risk of primary cardiovascular events (ATPIII model, current PCE model, SCORE, AGLA and current secondary prevention schema), machine learning models applied to analysis of prospective cohorts and electronic medical records data. Epidemiological aspects of analysis: confounding, biases and imprecision specific to prospective cohorts, electronic medical records data, biobank data. 2. Some data preparation will be required, coordinated with data preparation and harmonization done at BWH, Boston, MA. 3. Develop a machine learning model of secondary ASCVD. 4. Compare its performance to classical risk prediction model developed in a separate project.
**Grading:** - To receive a 6 the student must: meet the general goals and timeline of the thesis; work independently, demonstrate curiosity in the thesis topic and contribute her/his own ideas; communicate intermediate results clearly; write detailed and clear intermediate and final reports; document well project code; give a clear final presentation; work is exceptional and at a similar level expected for acceptance at leading international conferences or peer-reviewed journals. - To receive a 5 the student must: meet the general goals of the thesis; work independently; write a detailed and clear final report, and give a clear final presentation; document well project code. - To receive a 4 the student must: partially meet the general goals of the thesis; work somewhat independently; write a satisfactory final report, and give an understandable final presentation; document well project code
**Supervisor:** Prof. Dr. Gunnar Rätsch D-INFK **Advisors:** Dr. Olga Demler, D-INFK/ Brigham and Women’s Hospital, Harvard Medical School Dr. Samia Mora, Brigham and Women’s Hospital, Harvard Medical School Olga Mineeva, D-INFK
The goal of this collaborative project with colleagues from Brigham and Women’s Hospital/Harvard Medical School is to develop and validate a novel risk score (Residual Risk Score (RRS)) for predicting absolute risk of future events among patients with ASCVD using machine learning models and compare its performance to traditional risk prediction models. This project will contribute clinically meaningful and readily applicable results to directly impact a common clinical problem with high morbidity and cost burden; namely, reducing residual risk in ASCVD patients and personalizing their treatments. The RRS will be provided as a free clinical tool and could be incorporated in the electronic health record or as an application. The RRS would identify higher risk ASCVD patients more likely to suffer a recurrent event, and could be used to tailor novel and costly risk reduction strategies to higher risk patients.
The goal of this collaborative project with colleagues from Brigham and Women’s Hospital/Harvard Medical School is to develop and validate a novel risk score (Residual Risk Score (RRS)) for predicting absolute risk of future events among patients with ASCVD using machine learning models and compare its performance to traditional risk prediction models. This project will contribute clinically meaningful and readily applicable results to directly impact a common clinical problem with high morbidity and cost burden; namely, reducing residual risk in ASCVD patients and personalizing their treatments. The RRS will be provided as a free clinical tool and could be incorporated in the electronic health record or as an application. The RRS would identify higher risk ASCVD patients more likely to suffer a recurrent event, and could be used to tailor novel and costly risk reduction strategies to higher risk patients.
Please contact Olga Mineeva at olga.mineeva@inf.ethz.ch for further information on this project.
Please contact Olga Mineeva at olga.mineeva@inf.ethz.ch for further information on this project.