This study retrospectively analyzed real-world longitudinal data from 10 community health centers in Nanjing, China. The Nanjing source database covered the period 2015-2024, whereas individual patients had variable follow-up durations (Figure 1). Eligible patients were adults aged over 18 years with confirmed type 2 diabetes according to International Classification of Diseases, 10th Revision (ICD-10) codes and enrollment in the public health management program. Patients with severe cardiovascular or cerebrovascular diseases and pregnant women were excluded. A total of 50,288 individuals with type 2 diabetes were identified in the source database. After excluding individuals without analyzable longitudinal follow-up because of relocation, transfer out of the local public health management system, or incomplete or invalid records, 41,398 patients were included in the analytic cohort. Deaths observed after cohort entry were retained as outcome events and modeled as the absorbing state S5 rather than treated as loss to follow-up. The analytic cohort had a mean follow-up duration of 4.2 years and comprised 224 source data fields and 174,041 longitudinal follow-up records after analytic restructuring. Before risk-model development, identifiers, dates, per-measurement metadata, administrative and follow-up-process variables, outcome and survival-time variables, and fields used solely to construct MDP states and transitions were excluded from candidate prediction features. Eligible variables were subsequently encoded into 124 candidate design predictors as described below.

Patients were categorized into standardized and nonstandardized follow-up groups based on their observed follow-up patterns. Standardized follow-up was defined as 4 annual in-person visits, 1 per quarter, as stipulated by Chinese public health regulations. Nonstandardized follow-up was defined as any observed pattern that deviated from this guideline-based schedule, including fewer or more than 4 annual follow-ups, uneven follow-up intervals, or the use of telephone or community-based follow-up modalities instead of the scheduled in-person visits. The standardized follow-up group included 20,721 patients and the nonstandardized follow-up group included 20,677 patients. Because of the large sample size, several baseline variables showed statistically significant differences between the 2 groups; therefore, standardized mean differences were used as the primary criterion for assessing baseline comparability. Most measured baseline characteristics showed small between-group differences. No additional propensity score matching or weighting was performed because MDP transition-probability estimation required preserving the full observed state-action transition structure and avoiding sparse transition cells. The nonstandardized follow-up group was not a randomized intervention arm but an observational category derived from routine follow-up records. The comparison between standardized and nonstandardized follow-up patterns was treated as a descriptive assessment of observed follow-up patterns rather than as a causal estimate of treatment effect (Table 1).

External evaluation of the complementary risk-prediction models used an independent cohort of 25,506 patients from 10 communities in Shanghai, with a mean follow-up duration of 1.87 years (data collected in 2023-2024). Variables required for the primary risk models were harmonized in definition and coding across the Nanjing and Shanghai cohorts. The Shanghai cohort was used only to evaluate the external discrimination of the Nanjing-developed risk models and was not used to validate the MDP policy or cost-effectiveness projections.

This study was approved by the Ethics Board of Nanjing University, Nanjing, China (approval number OAP20240930001), and the approval covered the retrospective analysis of both the Nanjing development cohort and the Shanghai external validation cohort. The requirement for informed consent was waived because all data were deidentified and collected during routine clinical practice and public health management. No compensation was provided to participants, and no identifiable individual information is presented in this paper or its multimedia appendix.

We formulated personalized follow-up planning as a finite-horizon MDP with 12 monthly decision cycles. The model consisted of health states, follow-up actions, action-specific transition probabilities, and an immediate monthly net health benefit reward [20,21]. Table 2 summarizes the notation, state definitions, decision variables, health utility parameters, cost parameters, willingness-to-pay thresholds, and travel-related disutility parameters used in the model.

Health states were operationally defined using structured diagnosis records coded according to the ICD-10 Healthcare Security Version [22] issued by the National Healthcare Security Administration of China. Diabetes-related complication states were identified using ICD-10 code categories corresponding to type 2 diabetes with renal complications, ophthalmic complications, neurological complications, peripheral circulatory complications, and diabetic foot. The health state space consisted of 6 mutually exclusive states: no recorded diabetes-related complication (S0), diabetic nephropathy (S1), diabetic retinopathy (S2), diabetic peripheral neuropathy (S3), lower-extremity arterial disease and diabetic foot (S4), and death (S5; Figure 2). Baseline prevalent complications were defined as relevant ICD-10–coded diagnoses recorded before or at model entry, whereas incident complications were defined as the first newly recorded ICD-10–coded diagnosis of a complication category during follow-up among patients without that complication category at baseline. Death was identified from mortality records and modeled as the absorbing state. Once a patient entered a complication state, reversal to S0 was not allowed in the model. The detailed ICD-10 Healthcare Security Version coding algorithm and state-assignment rules are provided in Tables S19 and S21-S23 in Multimedia Appendix 1.

Actions were the decisions or measures that could be taken in each state and that influenced state transitions within the model. In the model constructed in this study, decisions involved determining the number of follow-ups, intervals, and methods for a patient over the course of a year. Periods were defined as t=1, 2, ..., 12, corresponding to the months of the year. Clinical events and follow-up records with irregular dates were assigned to the corresponding monthly cycle according to their recorded dates. Monthly state-action transitions were then constructed from the patient’s recorded health state and follow-up action within each monthly cycle.

Let a_t denote the follow-up action selected in month t, where a_t=0 represents no follow-up, a_t=1 represents outpatient follow-up, a_t=2 represents telephone follow-up, and a_t=3 represents community follow-up. Let c⁰_t denote the cumulative number of outpatient follow-up visits completed before month t.

Rewards were defined as immediate monthly net health benefits rather than as ratios of health benefits to costs. The MDP objective was to maximize the expected cumulative net health benefit over the 12-month horizon. Monthly health benefits were measured in quality-adjusted life days (QALDs), and monthly costs were converted into QALD-equivalent losses using the willingness-to-pay threshold.

For patient i in month t, current state s, action a, and prior follow-up count c⁰_t, the adjusted health utility was defined as follows:

For the incident-complication model, after excluding 669 patients with prevalent complications, removing 19,144 postevent person-years, and applying the prespecified outcome-time eligibility criteria, 152,415 eligible patient-year observations from 40,729 patients were included, with 7050 first incident complication events. After prespecified exclusions and one-hot encoding, 124 candidate design predictors entered the LASSO-Cox model; the lambda.1se rule retained only duration of diabetes. The apparent, optimism-corrected, and Shanghai external C-indices were 0.777, 0.777, and 0.778, respectively. The mortality model included 174,041 eligible patient-year observations from 41,398 patients and 3047 deaths. Of the same 124 candidate design predictors, 7 were retained by lambda.1se; the corresponding apparent, optimism-corrected, and external C-indices were 0.768, 0.767, and 0.812. Selected predictors are shown in Table 5.

In the prespecified lambda.min secondary analysis, 65 predictors were retained for incident complications and 95 for mortality. External C-indices were 0.752 and 0.833, respectively, compared with 0.778 and 0.812 for the primary lambda.1se models. Four selected categorical variables (occupation, marital status, medical payment type, and blood group) were structurally unavailable in Shanghai and, for the lambda.min external-discrimination analyses only, were fixed at their development-cohort reference levels. Predictor imputation was completed separately within the Nanjing development cohort and the Shanghai external-validation cohort before model fitting, and no additional imputation was performed during model fitting or external scoring. Neither primary lambda.1se model included these variables, so primary external validation was unaffected. In single-year cross-sectional analyses, duration of diabetes alone yielded C-indices of 0.790 in the second follow-up year (32,923 patients; 1194 events) and 0.768 in the third follow-up year (25,564 patients; 1383 events). These secondary analyses are reported in Table S24 in Multimedia Appendix 1.

Scaled Schoenfeld residual tests showed no evidence of violation of the proportional hazards assumption for the incident-complication model (global P=.93), whereas significant deviations were detected for the mortality model (global P<.001). Therefore, we used the LASSO-Cox models primarily as prediction and risk-stratification tools rather than as causal explanatory models. To examine whether relaxing the proportional hazards assumption materially affected model discrimination, we conducted a sensitivity analysis using random survival forest models that do not rely on the proportional hazards assumption. For the incident-complication model, the 500-tree random survival forest achieved a development C-index of 0.781 and an external C-index of 0.776, compared with 0.777 and 0.778 for the LASSO-Cox model, respectively. For the mortality model, the random survival forest achieved a development C-index of 0.791 and an external C-index of 0.814, compared with 0.768 and 0.812 for the LASSO-Cox model, respectively. The corresponding external C-index differences relative to LASSO-Cox were −0.002 for incident complications and +0.002 for mortality. These differences were negligible in magnitude, indicating that the random survival forest did not materially improve external discrimination. For mortality, the detected proportional-hazards deviations did not materially weaken discrimination; for incident complications, the random survival forest served as a nonparametric sensitivity check despite no detected proportional-hazards violation. We retained the LASSO-Cox models as the primary risk models because they provide a parsimonious and interpretable set of predictors. Detailed results of the random survival forest sensitivity analysis are provided in Table S17 in Multimedia Appendix 1.

The MDP model generated a 12-month personalized follow-up policy for each example patient according to the patient’s initial health state, age, and travel distance. Clinical variables from electronic health records were used for state identification and in the risk prediction models. The MDP itself used the current health state, age, monthly health utility, travel distance, and smoothed transition probability matrices to determine the optimal follow-up schedule and modality.

Table 6 summarizes the MDP-generated personalized follow-up strategies for example patients entering the model from different initial health states. The table reports the annual number of follow-ups and the distribution of follow-up modalities; the complete month-by-month schedules are provided in Table S1 in Multimedia Appendix 1. Patients in S5 were assigned no follow-up because death was modeled as an absorbing state. If a patient’s complication status changed during follow-up, the subsequent strategy would be updated according to the new health state. The illustrative S0 patient was aged 60 years and had a travel distance of 2.0 km. Group-level expected follow-up counts are reported separately in Table S11 in Multimedia Appendix 1.

At the individual level, the personalized follow-up strategy improved expected health outcomes for the example patient. For the example patient initially in S0, the personalized strategy increased annual QALDs from 313.64 to 314.12, corresponding to an increase of 0.48 QALDs per year. A Monte Carlo algorithmic check with 10,000 simulated trajectories produced similar results: the personalized strategy yielded 314.26 QALDs compared with 313.75 QALDs under standardized follow-up. In this check, the final complication-state probability was 5.19% under the personalized strategy and 5.09% under the standardized strategy, and the corresponding final death probabilities were 1.28% and 1.61%. These Monte Carlo estimates are reported in Table S4 in Multimedia Appendix 1. Complete expected-value effectiveness results across all initial states are provided in Table S2 in Multimedia Appendix 1.

Annual costs were obtained by accumulating monthly state and follow-up costs over the 12-month horizon. For the example patient initially in S0, the personalized strategy increased effectiveness by 0.48 QALDs and increased annual cost by 33.46 CNY. The resulting ICER was 25,499 CNY/QALY, or 69.86 CNY/QALD. The Monte Carlo algorithmic check produced broadly consistent cost-effectiveness results for the S0 example patient. The willingness-to-pay threshold, set at 3 times China’s 2024 per-capita gross domestic product, was 287,248 CNY/QALY. The ICER was substantially below this threshold, indicating that the personalized follow-up strategy was cost-effective for the example patient under the model assumptions, although it was not cost-saving under the base-case program-costing assumptions. The complete individual-level cost-effectiveness results across initial health states are provided in Table S3 in Multimedia Appendix 1.

At the group level, we evaluated the MDP-based personalized follow-up strategy in a heterogeneous simulated cohort of 1000 patients initialized in S0. The analysis used a fixed cycle duration of 365/12 days for each of the 12 monthly cycles. The personalized strategy produced a mean annual effectiveness of 314.05 QALDs (0.86041 QALYs), compared with 313.52 QALDs (0.85896 QALYs) under the standardized strategy. The mean incremental effectiveness was 0.529 QALDs (0.00145 QALYs) per patient-year. The cohort-mean month-by-month accumulation of QALDs and costs under both strategies is reported in Table S10 in Multimedia Appendix 1; the 12 monthly values sum exactly to the annual cohort means reported in Tables S5 and S6 in Multimedia Appendix 1. The corresponding annual reconciliation based on patient-level and monthly sums is provided in Table S7 in Multimedia Appendix 1.

The cohort-level QALDs were consistent with the model’s state utilities and end-of-horizon state distributions. For reference, a patient remaining in S0 throughout the full year without outpatient travel disutility would accrue 365 × 0.868 = 316.82 QALDs. The slightly lower cohort means reflected transitions to lower-utility complication states, mortality, and outpatient travel disutility.

We also compared monthly expected state distributions under the 2 strategies, including complication-state occupancy and cumulative mortality, as reported in Tables S12 and S13 in Multimedia Appendix 1. By the end of the 12-month horizon, the expected proportion of patients occupying complication states S1-S4 was 5.114% under the personalized strategy and 5.115% under the standardized strategy, corresponding to 51.14 and 51.15 expected patients per 1000 patients, respectively. The corresponding expected cumulative mortality proportions were 1.437% and 1.610%, corresponding to 14.37 and 16.10 expected deaths, respectively. The 2 strategies therefore produced broadly comparable end-of-horizon state distributions, and the group-level advantage of the personalized strategy lay in improved QALDs and cost-effectiveness rather than in a large reduction in end-of-horizon complication or death probabilities. The personalized strategy nevertheless resulted in slightly lower cumulative mortality and 0.0126 additional expected living patient-months per patient over the 12-month horizon.

In the present group-level simulation, patients were initialized in S0 to evaluate personalized follow-up planning among patients without baseline complications. The base utility of S0 was fixed at 0.868 for all patients, whereas age, travel distance, and action-specific travel disutility contributed to heterogeneity in simulated outcomes. The personalized strategy generated a mean of 2.72 follow-ups per patient-year, including 1.74 outpatient follow-ups, 0.96 telephone follow-ups, and 0.01 community follow-ups; the standardized strategy generated 3.98 outpatient follow-ups per patient-year. The corresponding group-level strategy summary is provided in Table S11 in Multimedia Appendix 1.

Under the specified program-costing assumptions, the personalized strategy increased mean annual cost by 24.67 CNY and mean annual effectiveness by 0.00145 QALYs compared with standardized follow-up. The resulting ICER was 17,016 CNY/QALY, which was substantially below the willingness-to-pay threshold of 287,248 CNY/QALY. The complete group-level cost-effectiveness results are provided in Table S5 in Multimedia Appendix 1.

To examine whether the MDP results were sensitive to the smoothing assumptions used in transition probability estimation, we repeated the group-level MDP analysis using alternative Laplace smoothing parameters of α=.1, .5, 1.0, and 2.0; detailed results are provided in Table S14 in Multimedia Appendix 1. Across these smoothing parameters, the personalized strategy increased mean annual effectiveness by 0.00136-0.00178 QALYs and increased mean annual cost by 19.95-53.67 CNY compared with standardized follow-up. The corresponding ICERs ranged from 14,716 to 31,151 CNY/QALY. All ICERs remained substantially below the willingness-to-pay threshold of 287,248 CNY/QALY, indicating that the projected cost-effectiveness results remained consistent across the examined transition-probability smoothing parameters.

We also performed deterministic 1-way sensitivity analyses for health utility values, complication-related medical costs, the annual baseline-management cost ceiling c_max, follow-up costs, the algorithmic discount factor, and travel-related disutility. The parameter ranges are reported in Table S15 in Multimedia Appendix 1, and the corresponding results are presented in Table S16 in Multimedia Appendix 1. Across these deterministic scenarios, the personalized strategy increased annual effectiveness by 0.00124-0.00160 QALYs and increased annual cost by 21.74-55.68 CNY compared with standardized follow-up. The corresponding ICERs ranged from 14,987 to 35,613 CNY/QALY. The ICER for the personalized strategy remained below the willingness-to-pay threshold of 287,248 CNY/QALY across all examined deterministic scenarios. Varying the algorithmic discount factor only had a little effect on the conclusion: with no discounting, the personalized strategy increased annual effectiveness by 0.00145 QALYs and annual cost by 25.42 CNY, resulting in an ICER of 17,494 CNY/QALY. With a monthly factor of 0.9959, corresponding to an annual discount rate of 5%, the corresponding values were 0.00145 QALYs, 25.29 CNY, and 17,408 CNY/QALY. Both ICERs remained far below the willingness-to-pay threshold. In the willingness-to-pay–dependent policy re-optimization analyses, the MDP was re-solved using willingness-to-pay objectives corresponding to 1 and 3 times China’s 2024 per capita gross domestic product. Under the 1× gross domestic product objective, the personalized strategy increased effectiveness by 0.00140 QALYs and cost by 18.65 CNY, resulting in an ICER of 13,295 CNY/QALY. Under the 3× gross domestic product objective, the personalized strategy increased effectiveness by 0.00145 QALYs and cost by 24.67 CNY, resulting in an ICER of 17,016 CNY/QALY. Both ICERs were below their corresponding willingness-to-pay thresholds. Varying the annual baseline-management cost ceiling c_max to 12,000, 15,000, and 18,000 CNY/year produced ICERs of 14,987, 17,016, and 19,018 CNY/QALY, respectively; the 15,000 CNY/year scenario reproduced the base-case result.

The final risk-prediction models developed in Nanjing were applied without reestimating model parameters in the independent Shanghai cohort of 25,506 patients from 10 communities (mean follow-up 1.87 years). The eligible incident-complication validation set contained 25,499 observations with 1742 events and yielded an external C-index of 0.778. Mortality validation used all 25,506 observations, including 125 deaths, and yielded an external C-index of 0.812. As the Shanghai cohort had a shorter follow-up duration than the Nanjing development cohort, these analyses validate discrimination rather than the MDP policy, cost-effectiveness projections, or long-term clinical effectiveness.

Using longitudinal primary-care data, we developed a 12-month finite-horizon MDP that selected follow-up timing and modality by maximizing expected net health benefit under prespecified utility, cost, access, and willingness-to-pay parameters. In both the example-patient and simulated-cohort analyses, the personalized policy produced small projected QALY gains at modest incremental program cost. In the 1000-patient S0 cohort, the ICER was 17,016 CNY/QALY, below the modeled willingness-to-pay threshold of 287,248 CNY/QALY. End-of-horizon complication-state occupancy was nearly identical between strategies, whereas modeled mortality was slightly lower under the personalized policy. As action-conditional transitions were estimated from observational records and may reflect confounding by indication, these differences are model-based projections and should not be interpreted as causal effects of follow-up modality.

Deterministic sensitivity analyses preserved the overall cost-effectiveness conclusion across the examined transition-smoothing, utility, cost, travel-disutility, willingness-to-pay, cycle-convention, and discount-factor scenarios. These analyses support structural robustness within the specified model but do not quantify sampling or joint parameter uncertainty; accordingly, they do not provide a probability that the personalized strategy is cost-effective. The higher modeled cost arose from both additional modality-specific follow-up charges and greater accrual of baseline management costs over the slightly larger number of expected living patient-months.

Prior diabetes follow-up research has examined encounter frequency, guideline adherence, telemedicine, personalized care systems, and self-management support [15-17,33-36]. Risk-based follow-up has also been evaluated in other chronic-disease settings [18,37,38]. These studies inform who may need more intensive care or how follow-up can be delivered, but fewer provide an explicit sequential decision framework that jointly selects timing and modality while incorporating expected health benefit, accessibility, and resource use. Related methodological work has applied machine-learning methods to predict follow-up intervals and has used MDP or approximate dynamic programming formulations to schedule health care encounters and allocate resources in other service settings [20,21,39].

Our study extends this literature by formulating follow-up planning as a sequential decision problem and incorporating QALDs and program costs directly into the MDP reward structure. The model did not project short-term cost savings; instead, it generated a small additional expected health benefit at an incremental modeled cost. This distinction is important in primary care, where decision support should make the trade-off between expected benefit and resource use explicit rather than equating personalization with lower expenditure.

The framework should be viewed as a candidate decision-support approach rather than a ready-to-deploy scheduling system. After prospective validation, it could support structured planning of follow-up frequency and modality using routinely collected clinical information together with prespecified utility, cost, and accessibility parameters. Its effects on clinician workload, patient adherence, service substitution, and implementation feasibility were not evaluated here and should be established before the model is used to modify mandated or routine follow-up schedules.

The incident-complication model was highly parsimonious: lambda.1se retained duration of diabetes alone, with apparent, optimism-corrected, and external C-indices of 0.777, 0.777, and 0.778, respectively. The lambda.min model retained 65 predictors but had lower external discrimination, and the single-year analyses yielded similar discrimination with duration of diabetes alone. Several conventional metabolic markers, including fasting blood glucose, blood pressure, and BMI, provided little incremental 1-year discrimination beyond duration of diabetes in this analysis. This parsimony reflects both the 1-year prediction horizon and the prespecified lambda.1se tuning rule, under which duration of diabetes alone achieved discrimination comparable to that of the larger lambda.min model. These findings are specific to 1-year risk discrimination in the available data and do not imply that these or other established clinical risk factors are clinically unimportant. For mortality, lambda.1se retained 7 predictors; although lambda.min yielded a higher external C-index (0.833 vs 0.812), the prespecified parsimonious model remained primary rather than selecting a model based on external performance. The 500-tree random survival forest sensitivity analysis changed external C-indices by only −0.002 for incident complications and +0.002 for mortality. Thus, the detected proportional-hazards deviations in the mortality model did not materially impair discrimination. These risk models were complementary validation analyses and supplied no numerical input to the MDP.

This study has several limitations. First, action-specific transition probabilities were estimated from observational follow-up records rather than randomized or prospectively assigned interventions. They may therefore reflect care-seeking behavior, clinician decision-making, access, adherence, residual baseline imbalance, and unmeasured confounding; Laplace smoothing addresses sparse cells but not confounding by indication. Most baseline differences were small, although exercise frequency and blood type had standardized mean differences of .199. Transition-cell denominators represent patient-month transitions rather than distinct independent patients, and sampling uncertainty in these probabilities was not modeled. The incremental effectiveness estimate was also small relative to its variation across smoothing scenarios. The MDP results should therefore be interpreted as projected decision-support outputs under the estimated transition structure, rather than as causal evidence for reducing guideline-based follow-up requirements. Individualization was limited to health state, travel distance, and age because the LASSO-Cox risk estimates did not condition the action-specific transitions.

Second, random-forest–based predictor imputation does not propagate imputation uncertainty in the same way as multiple imputation. In addition, the Cox models were fitted to person-year records, so patients with longer follow-up contributed more records, although cross-validation, bootstrap resampling, and robust SEs were handled at the patient level. The models were therefore used for prediction and risk stratification rather than coefficient-level causal inference. The reported bootstrap optimism correction was conditional on the LASSO-selected predictor set and therefore quantified optimism in the Cox refitting step rather than uncertainty from repeating the complete variable-selection procedure. External calibration was not assessed, so the reported C-indices establish discrimination but not calibration of absolute risk.

Third, routine public health records did not capture medication costs, productivity losses, or the full opportunity cost of provider time, facilities, consumables, and administrative support. The analysis therefore used a restricted public health program perspective. The absence of an additional charge for the first 4 outpatient visits is a program accounting convention, not evidence of zero resource use. Adding omitted resource components could change the absolute costs and ICER in either direction, depending on differential resource use. As a substantial component of the incremental cost arises from baseline-management expenditure accruing over additional living patient-months, the cost-effectiveness ratio implied by this survival channel alone is close to the baseline-management cost divided by the utility of S0; its magnitude is therefore largely determined by this prespecified cost parameter. The baseline-management cost function was also prespecified rather than estimated from individual-level expenditure data; varying its annual ceiling, c_max, from 12,000 to 18,000 CNY/year assessed structural sensitivity but did not substitute for empirical cost validation.

Fourth, travel distance was represented as a linear accessibility-related disutility. This simplification may not capture nonlinear relationships or nontravel-related barriers and may favor lower-burden modalities for patients living farther from primary care. Part of the projected QALY gain therefore reflects modeled access burden rather than biomedical disease control alone.

Fifth, the MDP states were mutually exclusive and did not explicitly represent concurrent diabetes-related complications, which may underestimate the burden of multimorbidity. The 12-month horizon also captures short-term projected cost-effectiveness rather than lifetime value, as major diabetes-related complications often develop over multiple years.

Finally, the Shanghai cohort had shorter follow-up than the Nanjing development cohort and was used only to validate risk-model discrimination; neither external calibration nor the MDP policy or economic projections were validated in Shanghai. Prospective studies should evaluate implementation effectiveness, workload, adherence, modality-specific outcomes, calibration, and generalizability, while considering risk-adjusted or causal methods for estimating action-conditional transitions.

This study developed a 12-month MDP framework that uses observed action-conditional transition patterns together with prespecified utility, cost, access, and willingness-to-pay parameters to generate personalized follow-up policies for patients with type 2 diabetes. In the modeled analyses, the personalized strategy produced small projected QALY gains at modest incremental program cost, and independent Shanghai data supported the discrimination of the complementary risk-prediction models. As the decision model relied on observational transition data and partly prespecified economic parameters, these results should be regarded as model-based projections requiring prospective implementation, calibration, and economic validation rather than as causal evidence for changing guideline-mandated follow-up schedules.