People on long-term dialysis develop kidney cancer far more often than the general population. Japanese nationwide surveys put the excess at 12- to 40-fold and the annual incidence at about 119 cases per 100,000 patients (Ishikawa 1989). Most of these tumors grow inside the patient's own failed kidneys after those kidneys fill with fluid-filled cysts, a condition called acquired cystic kidney disease. Because the risk appears to climb with years on dialysis (52.3% of cases occurred after more than ten years of treatment, Ishikawa 2002), routine imaging of long-term patients has become the accepted response. The trouble is that years on dialysis and chronological age rise together. The dialysis population of that era was aging and surviving longer, so the evidence cannot tell us whether the dialysis environment speeds cancer formation or whether long-term patients are simply older. The survival evidence for screening is weak for a similar reason. It is measured from the date of diagnosis, so finding a tumor earlier makes survival look longer even if no patient lives any longer. In a population where heart disease kills most patients first, that distortion can be large.
This proposal is a retrospective cohort study that fixes both problems, and it rests on one claim. Screening eligibility for dialysis patients should be based on the excess cancer risk that remains attributable to dialysis vintage (years on dialysis) after attained age and sex are normalized. The value of that screening should be judged by kidney-cancer death counted from the moment a patient becomes eligible, not from diagnosis. The concrete output is a vintage-specific, age-normalized incidence curve and a head-to-head comparison of screening rules. The rules are triggered by vintage, by age, or by symptoms. They are compared at equal imaging budgets, with heart-related and other non-cancer deaths treated as a competing outcome, meaning an event that removes a patient before cancer can be observed.
Core Claim and What Is New
The Japanese surveys connect dialysis cancer risk, proactive imaging and retrospective follow-up (Ishikawa 2002, prognosis; Ishikawa 2005 review). They motivate this work but never perform it. They report case counts by vintage and age separately. They match surgical survivors to age-matched controls. They do not estimate incidence on a joint age-by-vintage grid with person-time denominators, and they do not align follow-up to a screening-eligibility start date. Demographic normalization has repeatedly shown that crude trends mislead. Digestive-disease cases rose 74.4% while age-standardized death rates fell 1.38% per year (Wang et al.). Age-adjusted rates exposed a hidden thyroid-cancer mortality gap among Filipino Americans (Hsu et al.). Marginal standardization separated a policy effect from a doubling of older mothers (Yan et al.). These papers supply the method, but none of them uses a second, disease-specific time axis such as dialysis vintage. The ARIES breast-screening study (Oberije, Glocker et al.) is the closest design across screening and demographics. It stratifies AI-assisted screening results by age and ethnicity in 306,839 mammograms. It asks whether a reader performs equally across subgroups, however, not who should be screened, and it does not involve dialysis.
What is new here is therefore not standardization itself. It is the use of normalization to answer a screening-eligibility question in a population with two clocks, attained age and vintage. The answer is then tied to an outcome that is protected from lead-time bias.
Research Program
The data requirement is a multi-center dialysis cohort with three components. It needs start dates of dialysis. It needs abdominal imaging events, with each one recorded as either a protocolized surveillance scan or a scan triggered by symptoms. It needs linkage to cancer and death records. The research file lists no public dataset of this kind. The ESRD hospital records in the AICare user study show that such EHR data exist at individual institutions, but they were assembled for one-year mortality prediction, not for cancer follow-up. Assembling the cohort is part of the work. Ideally it would span a center that screens routinely and one that does not, so that screening exposure varies.
The incidence model is a Poisson regression on person-time split into cells of attained age, sex, vintage and calendar period. Age and vintage each enter as smooth splines. Attained age equals age at dialysis start plus vintage, so the model includes age and vintage but not age at start. This avoids the exact linear dependence familiar from age-period-cohort analysis. The vintage effect is then identified by comparing patients of the same age who began dialysis at different ages. The fitted rates are directly standardized to a dialysis-internal reference population, namely the pooled person-time distribution across age, sex and vintage. The WHO standard is not used, because the dialysis age structure is atypical, an open problem the file flags. Results are re-reported under two alternative standards so the reader can see how much the choice matters. The excess over the general population is computed by indirect standardization against registry rates for the same age and sex. This replaces the crude 12–40-fold range with an age-normalized ratio that varies by vintage.
The screening evaluation copies the logic of a randomized trial that was never run, using the observational data. Each patient enters the analysis on the date they first meet a candidate eligibility rule. Patients are duplicated into each strategy whose rule they meet. Each copy is censored when the patient's actual care departs from that strategy, and the remaining copies are weighted by the inverse probability of staying on protocol. Those weights are modeled from comorbidity, cardiovascular history, age, cause of kidney failure and center. The primary outcome is the cumulative incidence of kidney-cancer death, with all other deaths treated as a competing event. Secondary outcomes are stage at detection and the number of nephrectomies for benign lesions. Because every strategy starts its clock at the same eligibility date, earlier detection cannot by itself lengthen measured survival.
Baselines
The naive reference point is the crude analysis implicit in the existing literature: case counts and crude incidence by survey wave, plus survival from diagnosis. The strongest existing quantitative result is the Ishikawa prognosis cohort of 848 patients (Ishikawa 2002). It reports overall 5-year survival of 64.0% and disease-specific 5-year survival of 81.5%. Among the 662 surgically treated patients, 5-year survival was 79.5%, against 78.1% for general dialysis controls. We will rerun that exact analysis on our cohort, as Kaplan–Meier from diagnosis comparing screen-detected with symptom-detected cases. This shows how much of any apparent screening benefit vanishes once time zero moves to eligibility. Beating this baseline does not mean a larger survival number. It means the estimate either holds or shrinks by a measured amount under the bias-protected design, with a confidence interval that excludes zero for the effect on kidney-cancer death.
The strongest methodological competitor is a standard Cox model with age entered as a linear covariate and vintage as a covariate, the approach a skeptical epidemiologist would default to. The propensity-adjusted design in Tewari et al., which found 68% lower cancer-specific death for prostatectomy than for conservative care after matching, represents the matching alternative for the screening comparison. Both models run on the same person-time, the same covariates and the same center adjustment. Two ablations isolate the key ingredient. The first drops the vintage term and keeps age only, which tests whether vintage carries risk once age is normalized. The second keeps diagnosis as time zero and treats competing deaths as ordinary censoring, which measures the combined size of lead-time bias and competing-risk bias.
Why not just…?
Why not just put age and vintage in a Cox model? This is a real threat. For estimating the vintage effect it is close to equivalent to our Poisson model, and we will report both. If a flexible-age Cox model gives the same vintage curve, standardization adds only a fixed reporting scale that can be compared across eras and countries. The analysis stands or falls on the joint age-vintage separation, not on the standardization step. What the experiment will show is whether a linear age term, the usual default, leaves residual age confounding in the vintage estimate. If it does, the spline-based version and the linear version will disagree.
Why not just keep screening everyone after ten years on dialysis? The data already rule out waiting for symptoms: only 5.8% of cases presented with symptoms in Ishikawa 2002, and metastatic tumors averaged 8.8 cm (Ishikawa 1989). The ten-year rule, however, comes from the share of cases occurring after ten years, which is a numerator statistic. It is not a rate threshold. If vintage risk is flat after age normalization, a rule based on age, or on cystic change seen on imaging, would find the same cancers with fewer scans. The budget-matched comparison decides this directly.
Why not just reanalyze the published Japanese tables? The published tables give marginal distributions by age and by vintage, not joint person-time. Reanalysis can provide a crude pilot bound but cannot separate the two clocks.
Why not just deploy a kidney-tumor AI model and screen more efficiently? The KiTS19 winning model reached a tumor Dice score of 0.851 (Heller et al.). Dice is a 0-to-1 measure of how well a predicted tumor outline overlaps the true one. That model was trained on contrast-enhanced CT, which is usually avoided in kidney failure. McGough et al. note that there are no labeled datasets of non-contrast or cyst-distorted kidneys. The file reports no result for any model on such kidneys, either from the McGough review or from a pathology-based model like Chen et al.. A better scan reader would also not answer the question of who should be scanned. The two efforts complement each other, and eligibility comes first.
Decisive Checks
The first check is whether the age-normalized incidence curve rises with vintage across patients of the same attained age, with confidence bands that exclude a flat curve. If it is flat, the vintage-triggered rule loses its rationale, and that negative result is publishable on its own. The second check is whether moving time zero to eligibility and accounting for competing deaths shrinks the apparent screening benefit, and by how much compared with the diagnosis-based survival comparison against the 79.5% versus 78.1% Ishikawa figure. The third check compares vintage-triggered, age-triggered and symptom-triggered rules at matched numbers of scans, in kidney-cancer deaths averted per thousand scans. Robustness will be tested across standard populations, across causes of kidney failure (glomerulonephritis was 77.4% of the 1988 cohort, against diabetic disease in many current cohorts), and with leave-one-center-out estimates, which address the geographic narrowness of the original evidence.
References
- Renal cell carcinoma in chronic hemodialysis patients in Japan
- Renal cell carcinoma in chronic hemodialysis patients. A questionnaire study in 2000.
- Prognosis in dialysis patients complicated with renal cell carcinoma.
- Present status of renal cell carcinoma in dialysis patients: A questionnaire study in 2004 and review of past questionnaires since 1982
- Global, regional, and national burden of digestive diseases: findings from the global burden of disease study 2019
- Characterization of mortality and high-risk characteristics of thyroid cancer in Filipinos using the California Cancer Registry
- The trend of caesarean birth rate changes in China after ‘universal two-child policy’ era: a population-based study in 2013–2018
- Assessing artificial intelligence in breast screening with stratified results on 306 839 mammograms across geographic regions, age, breast density and ethnicity: A Retrospective Investigation Evaluating Screening (ARIES) study
- Augmenting Clinical Decision-Making with an Interactive and Interpretable AI Copilot: A Real-World User Study with Clinicians in Nephrology and Obstetrics
- Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach
- The state of the art in kidney and kidney tumor segmentation in contrast-enhanced CT imaging: Results of the KiTS19 Challenge
- Artificial intelligence for early detection of renal cancer in computed tomography: A review
- Artificial intelligence for diagnosing and predicting survival of patients with renal cell carcinoma: Retrospective multi-center study