Five-Year Cost-Effectiveness Modeling of Primary Care-Based, Nonmydriatic Automated Retinal Image Analysis Screening Among Low-Income Patients With Diabetes

We constructed a decision tree to conduct a cost-effectiveness analysis comparing two screening strategies via Markov modeling with microsimulation. The experimental strategy involved ARIAS screening of all adult patients with diabetes performed in the primary care clinic, whereas the reference strategy was current standard of care: referring all patients with diabetes for a dilated screening eye examination.^12-14 Our model used customized data for our institution derived from a prospective study performed by our group to examine the effects of an ARIAS-based primary care clinic DR screen on patient adherence to ophthalmic follow-up recommendations.¹⁵ Briefly, 179 adults with diabetes were screened with ARIAS, performed using nonmydriatic (undilated) retinal photography and image grading with EyeArt 2.0 (Eyenuk, Inc., Woodland Hills, CA, USA; of note, EyeArt 2.0 recently received FDA clearance for clinic use of autonomous detection of DR, including vision-threatening DR [vtDR]). After screening using the system, most patients received their results immediately during the visit, including the severity of their DR, a picture of their fundus photograph, and recommended follow-up interval with an eye care specialist. Patients with referable DR (moderate DR or greater), vtDR (any central diabetic macular edema [DME] or severe DR or greater), or who had inconclusive screening results were referred for ophthalmic consultation. As is typical of many referrals primary care providers in our health care system make to specialists, a follow-up referral examination was scheduled by phone subsequent to the initial PCP encounter. Rates of adherence to these recommendations were then compared with historical adherence rates to ophthalmic consultation among adults with diabetes from the same clinic. We used the baseline demographic data, referral rates and adherence rates (Table 1), and DR severity distributions described in the referenced study for the present economic modeling analysis.¹⁵

We defined vtDR as the presence of any central diabetic macular edema (DME), presence of severe nonproliferative DR (NPDR), or presence of proliferative DR (PDR) (Figure 1). After adjusting published epidemiologic data to reflect the racial/ethnic demographics among patients at our institution, we estimated the prevalence of each possible vtDR subtype or “state” (Table 2) and used these estimates in the economic model.^16-19 As described in greater detail in the next section, treatment and complication rates during year one and years two to five for possible treatment approaches, as well as the rates of progression to severe vision loss (SVL) among patients who were treated, were based on the Diabetic Retinopathy Clinical Research Network (DRCR) Protocol I and S studies.^18-22 The rate of progression to SVL for patients with vtDR who did not receive treatment was modeled based on epidemiologic data. These figures can be found in Table 3.^18-23

Costs were drawn from Medicare 2019 allowable charges using current procedural terminology (CPT) codes and other cost estimates in the recently published literature, adjusted to reflect 2019 dollars.^12-14,25 Per-patient treatment and complication costs for year one were as outlined in the respective studies for the relevant vtDR pathology. Annual per-patient treatment and complication costs for years two to five were modeled as the average for all treatments/complications reported between years two to five in the respective protocols for the patient’s corresponding retinopathy. These figures attempted to capture the overall cost from the payor’s perspective for the diagnosis and treatment of different vtDR states. In rare instances where the data were unclear, most notably regarding the number of annual visits to an ophthalmologist for vtDR treatment and monitoring (ie, the number of patient visits in DRCR Protocols I and S likely overestimate those of actual clinical practice), appeal to conservative expert opinion was made. All variables used in the model are listed in Tables 3 and 4.^{14,18-22,24,26,27} For disease that was below the threshold for treatment (severe NPDR without DME), we performed modeling based on an assumption that such patients would be examined three times per year. For treatable disease, three approaches were modeled: (1) DME in no DR or any nonproliferative DR was treated per the deferred laser arm of DRCR Protocol I; (2) PDR without DME was treated per the PRP arm of DRCR Protocol S; and (3) concurrent PDR and DME were treated per the ranibizumab arm of DRCR Protocol S (Table 2). Per-patient treatment and complication costs among the three treatment approaches were calculated for year one and years two to five (Figure 2A and B). Future outcomes were discounted at a future rate of 3% to reflect their present value.

As ARIAS is primarily a secondary prevention strategy, all patients were modeled with good initial visual acuity of 20/30 or better in the better-seeing eye. SVL was defined as progression to visual acuity worse than 20/200 in the better-seeing eye, and rates of development of SVL in treated and untreated patients were based on the published literature as described above (Table 3).^23,28

The primary outcome measure was the incremental cost-utility ratio (ICUR) of current screening practice compared with ARIAS screening, a value generally used to quantify the incremental cost associated with a one-unit increase in utility of one health intervention compared to another.²⁹ Using Markov modeling with microsimulation over a five-year period, the ICUR was determined by dividing the difference in average direct per-patient costs between the two screening strategies by the difference in mean health utility values between the strategies. For each of the two screening strategies, secondary outcome measures included the predicted proportion of patients with vtDR who were adherent with eye examination referral, and incident development of SVL after five years.

Health utility is a measure of quality of life with a time component that is ultimately totaled to calculate quality-adjusted life years (QALY). Utility values used in our model were obtained from previous studies converting visual acuity to QALY.^12-14,28,30 When two sources reported different QALY approximations for the same visual acuity, the more conservative, that is, higher, utility value was chosen.^28,30 For health utility changes associated with treatment by either ranibizumab and/or pan-retinal photocoagulation (PRP), we used utility values previously published (Table 5).¹⁴

As the primary care clinic at our institution serves approximately 3000 patients with diabetes each year, the relevant per-patient ARIAS screening costs were calculated using this number (Table 6). Published long-term outcome data from Protocols I and S extended to year five, allowing us to use these data in our five-year model.^20,22 Furthermore, these data guided our analysis of the model using a two-stage Monte Carlo simulation consisting of 100 iterations of bootstrap sampling of 3000 individual patients modeled over five years, in each branch of the decision tree allowing a comparison of 3000 patients in the current practice branch to be compared with 3000 patients in the ARIAS branch. This resulted in a total of 600 000 simulated patient experiences (100 × 3000 × 2 decision tree branches). Treatment costs and utility for the two management strategies across each of the trials were summed to create means and SDs permitting calculation of the ICUR.

One-way sensitivity analyses were performed for key variables. Subsequently, the variables were substituted with distributions to facilitate probability sensitivity analyses. Use of distributions instead of variable point estimates allowed for exploration of variables that may be exerting excessive influence while accounting for changes in multiple variables (as compared with a one-way sensitivity analysis where all other variables remain constant). Variables were tested over a distribution range (Table 7). The willingness-to-pay, or the dollar amount that health care decision-makers would be willing to pay for a one-unit increase in effectiveness, was set at $100 000.^13,14

Modeling was performed using TreeAge Pro software (TreeAge Pro 2019, R1. TreeAge Software, Williamstown, MA, USA) using methods similar to those in previously published cost-effectiveness literature.^12-14 Additional statistical analysis was performed using STATA/IC software version 14.2 (StataCorp LP, College Station, TX, USA) with alpha set at 0.05.

We determined the absolute and incremental costs and utility between the current practice and ARIAS groups (Table 8). At five years, current practice showed comparable utility as ARIAS, but at a much greater cost. We calculated a 23.3% reduction in cost in the ARIAS group compared with standard practice (P < .001). Comparing current practice to ARIAS screening, we computed an ICUR of $258 721.81.

In one-way sensitivity analysis, the cost-effectiveness of ARIAS was robust across a wide range of variability in individual model inputs (Table 9). The model favored ARIAS until the price of 0.5-mg ranibizumab fell below $80.26 (not including the cost of the injection procedure). Additionally, ARIAS was cost-effective up to a combined cost of ARIAS screening of $161.14 per screening and up to the standard of care eye examination adherence rate of 71.5%. The full list of the nine variables that had a threshold value leading to a reversal of the cost-effectiveness analysis from favoring ARIAS to favoring current practice, in addition to select variables that notably did not have a threshold value, are included in Table 9. All other variables not included in Table 9 had no significant effects on cost-effectiveness.

Following one-way sensitivity testing, each variable was converted from a point estimate to a distribution in the Markov model. This permitted probabilistic sensitivity analysis which identified the proportion of iterations favoring ARIAS versus standard of care at varying willingness to pay thresholds. As illustrated in Figure 3, at a willingness to pay of $50 000, the model favored ARIAS in 66% of samples, whereas at a willingness to pay of $100 000, the model favored ARIAS in 59% of instances.

Similar to findings in other cost-effectiveness studies, the per-patient price of 0.5-mg ranibizumab accounts for the greatest proportion of the annual cost for any vtDR (Figure 2A).¹³ As illustrated in Figure 2B with the medication cost of 0.5-mg ranibizumab removed, the second largest per-patient cost was the expense of ongoing clinic visits (eye examination with optical coherence tomography).

Compared with current practice, our five-year decision tree simulation showed that implementation of ARIAS increased the proportion of all patients with vtDR who were adherent with eye examination referral from 33.8% in the reference group to 59.4% in the ARIAS group. In this model, ARIAS also decreased the incidence of SVL from 3.1% to 1.1% (Table 10). These proportions were calculated from the decision tree in a deterministic fashion, precluding the use of “random walk” simulations, as we performed for primary outcome measures. Subsequently, we could not conduct statistical testing or sensitivity analyses to further analyze these estimates.