Summary
An analysis of All of Us health records estimated that about one in nine adults who had experienced Long COVID carried a corresponding diagnosis code, under assumptions comparing the records with federal survey data. Models rebuilt from released counts showed moderate ability to predict later coding, but performance varied by metric and site.
A medRxiv preprint analysing 23,790 adults in the US All of Us research programme estimates that about one in nine adults who had experienced Long COVID carried a corresponding electronic health record (EHR) code. The estimate depends on comparing the study records with federal survey prevalence under assumptions that make the two sources comparable.
The researchers also tested whether models built outside the All of Us research enclave from released counts could predict who would later receive a Long COVID code. Those models showed moderate discrimination, but their performance differed across evaluation measures, and their predicted risks were too high at held-out sites.
Comparing records with survey estimates
The authors compared the share of participants with a Long COVID code against prevalence estimates from the Household Pulse Survey. On that basis, they estimated that roughly one in nine adults who had ever had Long COVID carried its code. The authors also reported a women-to-men coding ratio of 0.79, with a 95% sensitivity interval of 0.65–0.95.
These are estimates of recorded diagnosis relative to survey prevalence, not a direct count of every person with the condition. EHR coding reflects what is entered into medical records; the comparison therefore relies on the assumption that the populations and measures are sufficiently comparable. The ratio indicates a difference in recorded coding, but the abstract does not identify its cause.
Predicting later diagnosis codes
For the prediction analysis, the researchers selected 23,300 adults who had no U09.9 code by day 90 after COVID-19 and predicted a first such code by day 455. U09.9 was the coded outcome: the model was predicting later documentation in an EHR, rather than independently assessing whether participants developed Long COVID.
The researchers trained naïve Bayes and linear discriminant models outside the enclave using released counts, then compared them with an elastic-net model fitted inside the enclave using the same patients and features. Evaluation used a random test set of 4,680 people, including 50 who received a code, and 18 held-out EHR sites with 5,760 people, including 60 coded cases.
The released models had areas under the receiver operating characteristic curve (AUROC) of 0.68 on the patient test set and 0.64 at held-out sites. The authors reported that these were 0.049 higher than the elastic-net model’s AUROC (95% confidence interval, 0.001–0.104). On the patient test set, however, the elastic net had a higher area under the precision-recall curve, by 0.062 (95% confidence interval, 0.003–0.122). The released models also overestimated coding risk at held-out sites.
AUROC and precision-recall curves capture different aspects of classification performance, so the results do not point to a consistent advantage across measures. The study’s central practical finding is that models rebuilt from released counts retained moderate ability to distinguish who would later receive a code, while the undercoding estimate highlights how much recorded diagnoses can differ from survey-based prevalence.
The report is a medRxiv preprint. Its authors say aggregate statistics, code, a notebook and the two released models are available through the project repository.