Abstract
Background: Image-derived artificial intelligence (AI) risk models have shown promise in short-term risk assessment for improving breast cancer screening. No image-derived long-term AI risk model for primary prevention has been developed and externally validated.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
Background: Image-derived artificial intelligence (AI) risk models have shown promise in short-term risk assessment for improving breast cancer screening. No image-derived long-term AI risk model for primary prevention has been developed and externally validated.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
A long-term image-derived AI risk model for primary prevention of breast cancer
Mikael Eriksson
Abstract
Background: Image-derived artificial intelligence (AI) risk models have shown promise in short-term risk assessment for improving breast cancer screening. No image-derived long-term AI risk model for primary prevention has been developed and externally validated.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
Background: Image-derived artificial intelligence (AI) risk models have shown promise in short-term risk assessment for improving breast cancer screening. No image-derived long-term AI risk model for primary prevention has been developed and externally validated.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
Methods: We performed a two-site case-cohort study of women aged 30-90 in a population-based screening study including two screening settings in Olmsted County, Minnesota (U.S.) and the KARMA cohort (Sweden) with women recruited between 2009-2017. Median follow-up was 10 years. An image-derived AI-risk model was developed in an independent Swedish population and we report on the validation in the Olmsted/KARMA studies. Absolute 10-year risks were calculated at study entry. Time-dependent Area Under the receiver operating characteristics Curve (AUC(t)) and the ratio of expected versus observed events (E/O) were estimated. Comparison with the clinical Tyrer-Cuzick v8 model was performed in KARMA using clinical guidelines. Analyses were performed for risk of all breast cancer and restricted to invasive cancer alone.
Results: The Olmsted/KARMA case-cohorts included 8,721 women with mean age 54.4 years (SD 10.6) in the subcohort and 1,633 incident breast cancer cases with mean age 57.0 years (SD 10.6). The image-derived AI 10-year average risk was estimated as 3.85% in Olmsted and 3.16% in KARMA. The E/O ratio was 1.01 (95% CI 0.95-1.06) in Olmsted and 0.98 (95%CI 0.90-1.07) in KARMA. The 10-year AUC(t) was 0.71 (95%CI 0.68-0.73) in Mayo and 0.72 (95%CI 0.69-0.77) in KARMA. Using the National Institute for Health and Care Excellence (NICE) guidelines, considering women at 8% as high risk, 32% of breast cancers could be subject to preventive strategies in the 9.7% of women at high 10-year risk based on the AI risk model. The corresponding numbers were 7.2% and 2.2% for Tyrer-Cuzick. Results were similar when restricted to invasive cancers only.
Conclusions: The 10-year image-derived AI-risk model showed good discriminatory performance and calibration in the two case-cohorts and, showed a significantly higher discriminatory performance than the clinical Tyrer-Cuzick v8 risk model in KARMA. The image-derived AI-risk model has the potential for clinical use in primary prevention and targets up to one third of breast cancers.
{{ help_message }}
{{filter}}