These data draw on multiple sources, including a claims-based analysis performed by McKinsey &
Company using Komodo Health insurance claims data and social determinants of health data from
McKinsey & Company’s Health Institute. Perkins School for the Blind is responsible for the
conclusions and recommendations of the research. Data included open and closed claims from
patients 0-22 years old at the time of their first medical claim during the period 2016-2022.
Because there is not a single defined ICD-10 diagnosis code for CVI, we identified patients with
diagnosed CVI using the following criteria: at least one ICD-10 code for disorders of visual
cortex (e.g., cortical blindness) and at least one ICD-10 code for another visual disorder. In
addition to identifying patients with diagnosed CVI, we built and applied a machine learning
model to also identify patients with probable undiagnosed or likely CVI.
The machine learning model was trained on patients with diagnosed CVI (‘true positives’) and on
patients identified not to have CVI (‘true negatives’). We classified patients as ‘true
negatives’ based on absence of certain diagnoses and patterns of care, for example, visits with
a pediatric neurologist and ophthalmologist in conjunction with absence of diagnosis of disorder
of visual cortex. The model used patient ICD-10 diagnosis information, specialist visits,
imaging procedures, age, geographic location, and state-level social determinants of health data
to predict whether patients were likely to have CVI or not. The model performed well in terms of
overall accuracy (79% F1 score) and had strong specificity (99%) but lower sensitivity (71%),
meaning the model was unlikely to classify someone as having CVI who did not, but may miss cases
of CVI. A relatively conservative prediction model aligned with the goal of confidence in
predicted cases of CVI. By examining both patients with diagnosed CVI and patients with likely
CVI in claims data, we can more comprehensively explore the total burden of disease and
associated factors.
Several limitations warrant consideration in our analysis. Firstly, we utilized a combination of
open and closed medical claims records. Closed claims include the complete medical billing data
from a specific payer, whereas open claims derive from various sources and may not encompass all
billing data for a patient. Consequently, while open claims enable us to include a larger number
of patients in our analyses, we may not have access to all their medical billing data or
diagnoses. Moreover, our model is unlikely to account for children lacking health insurance or
those primarily receiving school-based care, potentially leading to an underestimation of the
burden impact of CVI within these populations. Additionally, our approach is inclined to
generate conservative estimates due to the machine learning model being trained on more severe
cases of CVI. Lastly, data privacy restrictions prevent us from observing perinatal and related
codes in the dataset, thus hindering our ability to observe premature birth or perinatal
injuries.
Use this map to explore how CVI prevalence and social risk factors overlap
1. Use the slider to filter counties by CVI prevalence:
2. Turn map layers on and off to explore how risk factors overlap:
Poverty rate: Percentage of population living below poverty level
Access to transportation: The Transportation Access Score measures access to public transportation and vehicle ownership rate.
The score is represented on a 0 to 10 scale, with higher scores indicating worse access to public and private transportation.