Knowledge of the transitions between healthcare utilization states, such as virtual care, urgent care, or hospitalization, is important for the effectiveness of prediction models and public health planning strategies. Yet, most of the distributions for such transitions for respiratory diseases remain unknown. Using Kaiser Permanente electronic health records, we parameterized time-to-event transition distributions for COVID-19, Influenza, and RSV. This will allow for better informed prediction models, as well as enable us to answer questions about how people infected with respiratory diseases are moving through the healthcare system and who is accessing certain types of care.
The "healthcare utilization cascade."
For all three pathogens, we parameterized two types of transitions — the first event to occur following a starting event, as well as reaching each acuity threshold, or outcome or more severe for each outcome from a starting event. The first transition type can help answer questions about how people are moving through the healthcare cascade. The outcome or more severe transition type can inform prediction and forecast models by answering how many people will seek at least a certain level of care in the healthcare cascade.
The outcome or more severe analysis was further parameterized by certain demographic covariates, such as age, gender, and vaccination status. Some covariates, such as age and Charlson weight, have large effects on the probability of needing any or higher levels of care. Yet, no variables have a meaningful impact — on the order of around half a day — on the rate of seeking such care. We provide the rate estimates for Outpatient or worse and Inpatient or worse for reference, though under the expectations that the rates don’t differ enough to merit inclusion into most models.
The tables containing the parameterized distributions, probability of occurance, and time-to-event estimates (examples below) for COVID-19, Influenza, and RSV are provided in the Parameter_Tables folder. The methods and further discussion of these results will be provided in an upcoming paper (link will be provided here when published).
We used the flexsurv package in R, which allowed us to create parametric mixture models. We employed a competing risk framework for the first event models, as there are multiple states that an individual can move to, some of which might occur before others and prohibit them from occurring. This package is also useful because it allows for flexibility in distributions. Using AIC, we compared the models’ performance using different distributions to find the best fitting distribution for each transition of interest.
Graphical representation of inclusions criteria.
Above shows a figure describing our inclusion criteria for acute respiratory infection events. We looked between -7 and +30 days of a positive test to find events. From those events, we looked 20 days out to find the first event, if any, and 60 days out to find any events more severe. We also provide first event model parameters using a 60 day follow up period, but differnces between the two are small.
This extended criteria allowed us to restrict ARI healthcare events to being related/caused by COVID infections, while still including healthcare events that might take more than 20 days (like death) after the initial event.
Below are example outputs for the first event and outcome or worse analyses (along with the stratified estimates) for COVID-19.
There are two first event tables for each pathogen: a first_event_20 that shows the parameterized progressions using a follow-up period of 20 days, and first_event_60, which uses a 60 day follow-up period.
The “First Event” following a “Starting State” for COVID Infections:
For example, there is a 36.8% chance that the first event following symptom onset is an urgent care visit. The median time to these urgent care visits is 2.85 days following symptom onset.
The distribution of progressing to each outcome or more severe state from each starting state:
For example, there is a 48.3% chance that the of experiencing an urgent care visit or worse event following symptom onset of a COVID-19 infection. The median time to this outcome is 4.04 days following symptom onset.
All of the Event or more severe state outcomes were stratified by available demographic variables. Presented here are the probabilities of each outcome stratified by various covariates.
As discussed, time-to-event rates vary little across demographic groups (usually on the order of half a day). Therefore, only the stratified estimates of the location parameter and median time-to-event for Outpatient or worse and Inpatient or worse are presented.
As a reminder, here are the parameters and mean time-to-events for Outpatient or worse and Inpatient or worse for COVID-19 across all groups.
| Outpatient + Distribution | Outpatient + Parameters | Outpatient + Median Time To Event | Inpatient + Distribution | Inpatient + Parameters | Inpatient + Median Time To Event |
|---|---|---|---|---|---|
| Log normal | logmean = 1.43, logsd = 0.934 | 4.17 (4.09, 4.24) | Log normal | logmean = 1.93, logsd = 0.971 | 4.17 (4.09, 4.24) |
And here are the stratified estimates.
| Characteristic | Group | Outpatient + Location Parameter Estimate | Outpatient + Median Time To Event | Inpatient + Location Parameter | Inpatient + Median Time To Event |
|---|---|---|---|---|---|
| Age (%) | 0-17 | 1.196 | 3.31 (2.46, 4.32) | 1.846 | 6.34 (4.11, 10.47) |
| Age (%) | 18-49 | 1.384 | 3.99 (2.68, 6.12) | 1.866 | 6.46 (3.49, 12.48) |
| Age (%) | 50-59 | 1.401 | 4.06 (2.67, 6.04) | 1.979 | 7.23 (3.69, 13.84) |
| Age (%) | 60-69 | 1.443 | 4.23 (2.76, 6.28) | 1.987 | 7.3 (3.88, 13.9) |
| Age (%) | 70-79 | 1.504 | 4.5 (3.02, 6.66) | 1.968 | 7.15 (3.71, 13.56) |
| Age (%) | 80-89 | 1.508 | 4.52 (2.97, 6.62) | 1.868 | 6.48 (3.53, 12.63) |
| Age (%) | 90+ | 1.545 | 4.69 (3.11, 7.12) | 1.907 | 6.73 (3.53, 12.29) |
| GENDER (%) | Female | 1.459 | 4.3 (3.87, 4.79) | 1.938 | 6.94 (6.06, 7.93) |
| GENDER (%) | Male | 1.383 | 3.98 (3.45, 4.63) | 1.963 | 6.77 (5.6, 8.16) |
| Race/ethnicity (%) | Asian | 1.405 | 4.07 (0, Inf) | 1.966 | 7.14 (0, Inf) |
| Race/ethnicity (%) | Black | 1.480 | 4.39 (0, Inf) | 1.941 | 6.97 (0, Inf) |
| Race/ethnicity (%) | Hispanic | 1.381 | 3.98 (0, Inf) | 1.959 | 7.09 (0, Inf) |
| Race/ethnicity (%) | Multiple | 1.438 | 4.21 (0, Inf) | 2.005 | 7.43 (0, Inf) |
| Race/ethnicity (%) | Native Am Alaskan | 1.549 | 4.71 (0, Inf) | 1.962 | 7.11 (0, Inf) |
| Race/ethnicity (%) | Other | 1.352 | 3.87 (0, Inf) | 2.172 | 8.78 (0, Inf) |
| Race/ethnicity (%) | Pacific Islander | 1.376 | 3.96 (0, Inf) | 1.576 | 4.84 (0, Inf) |
| Race/ethnicity (%) | Unknown | 1.254 | 3.5 (0, Inf) | 1.842 | 6.31 (0, Inf) |
| Race/ethnicity (%) | White | 1.500 | 4.48 (0, Inf) | 1.889 | 6.61 (0, Inf) |
| COVID Vaccinations (%) | \<3 | 1.355 | 3.87 (3.55, 4.21) | 1.809 | 6.1 (5.59, 6.71) |
| COVID Vaccinations (%) | 0 | 1.342 | 3.83 (3.42, 4.31) | 1.862 | 6.44 (5.68, 7.31) |
| COVID Vaccinations (%) | 3+ | 1.460 | 4.31 (3.86, 4.87) | 1.957 | 7.07 (6.25, 8.04) |
| NDI_group (%) | \<-1 | 1.448 | 4.25 (4, 4.51) | 1.934 | 6.92 (6.39, 7.5) |
| NDI_group (%) | -1 | 1.464 | 4.32 (3.93, 4.7) | 1.899 | 6.68 (5.94, 7.44) |
| NDI_group (%) | 0 - 1 | 1.411 | 4.1 (3.72, 4.49) | 1.941 | 6.96 (6.24, 7.79) |
| NDI_group (%) | \>1 | 1.384 | 3.99 (3.64, 4.36) | 1.948 | 7.02 (6.29, 7.88) |
| Charlson_wt_group (%) | 0 | 1.314 | 3.72 (3.42, 4.02) | 1.784 | 5.95 (5.43, 6.58) |
| Charlson_wt_group (%) | 1-2 | 1.442 | 4.23 (3.75, 4.77) | 1.854 | 6.39 (5.64, 7.33) |
| Charlson_wt_group (%) | 3-5 | 1.550 | 4.71 (4.17, 5.28) | 1.939 | 6.95 (6.06, 7.94) |
| Charlson_wt_group (%) | 6+ | 1.643 | 5.17 (4.61, 5.82) | 2.014 | 7.5 (6.54, 8.58) |