Who Gets a Blood Test? Correcting Long-Term Lymphocyte Recovery Estimates After Radiotherapy

Radiotherapy kills circulating lymphocytes, the white blood cells that carry out immune surveillance. It does this even when the tumor is far from any immune organ, because the whole blood volume passes through the radiation beam many times over a multi-week course. This depletion is called radiation-induced lymphopenia. Low lymphocyte counts during and shortly after treatment are consistently linked to worse survival, and hospitals are beginning to adopt dose limits for the spleen based on these links. The limits currently proposed, however, are tuned to lymphocyte counts measured no later than about four months after treatment. In the largest gastric cancer cohort, 87.4% of patients were still lymphopenic at day 120, and the authors state that recovery beyond six months was simply unobserved (Clinical Considerations of Splenic Dose Constraints to Mitigate Radiation‐Induced Lymphopenia). The obvious next step is to pull later blood counts from the hospital record and extend the recovery curves to two or three years.

That step contains a trap, and this proposal addresses it. After active treatment ends, blood counts in the electronic record are not taken on a schedule. They are taken when someone is worried: a fever, an admission, suspected recurrence. Healthy survivors drift to infrequent visits, while declining patients are tested often. A naive long-term recovery curve built from whatever counts exist will therefore over-represent sick patients at late time points. It will make persistent lymphopenia look more common and more lethal than it is, and it may change which spleen doses appear safe. Our core claim is that the reason each blood test was ordered can be read from routine hospital data, and that it is the missing ingredient for honest long-term estimates. Each count can be labeled as a scheduled surveillance draw or a symptom-triggered draw from its encounter context (outpatient follow-up slot versus emergency, inpatient, or unscheduled visit). With that label, we can model the testing process explicitly and reweight the observed counts. The payoff is the first defensible estimate of lymphocyte recovery at 12, 24, and 36 months after radiotherapy, and a direct test of whether day-120 spleen constraints still predict long-term immune recovery once testing bias is removed.

What already exists and what is missing

Three retrospective single-institution studies connect radiation dose, lymphocyte kinetics, and survival. Together they motivate the proposal. In 96 gastric cancer patients, spleen $V_5$ (the spleen volume receiving at least 5 Gy) predicted the lymphocyte minimum, and the drop from baseline to minimum predicted overall and disease-free survival. Follow-up of counts, however, stopped at the immediate post-treatment check (The relationship between splenic dose and radiation-induced lymphopenia). The later 159-patient study extended counts to days 60, 90, and 120. It defined a recovery index as the day-120 count divided by baseline, and proposed spleen $V_5$ cutoffs of 180.6 cm³ for grade 4 lymphopenia and 272.2 cm³ for a 60.9% reduction in risk of death (Clinical Considerations of Splenic Dose Constraints to Mitigate Radiation‐Induced Lymphopenia). In 139 stage III lung cancer patients receiving chemoradiotherapy followed by anti-PD-1 consolidation, counts at three fixed landmarks predicted overall survival but not progression-free survival. Those landmarks were baseline, fraction 20, and one month post-treatment. A 39-patient flow cytometry subset showed faster CD8+ than CD4+ recovery (Effectiveness and mechanisms of lymphocytes at different time points in predicting consolidation immunotherapy following adaptive chemoradiotherapy in locally advanced non-small cell lung cancer). All three use fixed, early landmarks and treat observed counts as if they were missing at random. None extends past four months.

A second group of papers constrains the design by documenting the relevant biases. A surgical mortality study notes that follow-up in retrospective cohorts is informatively incomplete (90-day mortality prediction in elective visceral surgery using machine learning: a retrospective multicenter development, validation, and comparison study). A study of immunotherapy-induced colitis notes that laboratory tests in oncology are missing for clinical reasons rather than by chance (Risk factors for steroid-refractory in immune checkpoint inhibitor-induced colitis: a retrospective cohort study). Work on time-related biases shows how immortal time and lead time can manufacture apparent benefits (Reply: Time‐Related Biases in Observational Studies of GLP‐1 Agonists and Cancer Risk). Adjustment for measured variables cannot remove unmeasured confounding (Systemic Biases and Causal Limitations in Retrospective Studies).

A third group supplies candidate methods, and the proposal deliberately treats them as competitors rather than as its contribution. These are continuous-time survival models that take irregularly sampled measurement streams directly, namely CoxSig and CoxSig+ (Dynamic Survival Analysis with Controlled Latent States), and a deep state-space model of 12,695 chemotherapy patients. The state-space model found low lymphocytes together with high neutrophils to be an acute mortality signal under nivolumab and targeted agents (A deep state-space analysis framework for cancer patient latent state estimation and classification from EHR time-series data). Both handle irregular timing, but neither models why a measurement exists. Both learn from the sampling pattern as if it were signal. Neither was applied to radiation dosimetry.

The obvious idea at this intersection is "extend follow-up and fit a continuous-time model." This proposal differs from it in one specific way. It treats the blood-test ordering process as a separate, observable, biased mechanism, and it estimates recovery after correcting for that mechanism.

Research Program

The data come from extending two existing cohorts using records already held by the treating hospitals. The first is the postoperative gastric chemoradiotherapy cohort (2010–2020; the 96- and 159-patient reports appear to come from overlapping patients at the same institution, and we will deduplicate rather than pool them naively). The second is the Taizhou stage III lung cancer cohort (2018–2024). For every patient we will extract every complete blood count up to death or last contact, together with the dose-volume data already computed in the original studies (spleen $V_5$ through $V_{40}$ and maximum dose for gastric cancer; target and lung dosimetry for lung cancer). We will also record chemotherapy regimen and the start date of any progression or later treatment. The added work is labeling. Each count is tagged as scheduled or triggered, using its encounter type and whether a clinical event occurred within seven days. Two clinicians will adjudicate a random 10% sample to measure how reliable the labeling is.

The lung cohort contains a built-in validation window. During consolidation anti-PD-1 therapy, a blood count is drawn before every infusion on a fixed cycle regardless of how the patient feels. For that period, counts are close to protocol-driven. This gives us a stretch of near-complete, non-selective sampling against which any correction method can be graded.

The model has three linked parts. The first is a mixed-effects trajectory model for log lymphocyte count, with patient-specific splines in time since radiotherapy end and dosimetric and baseline covariates as fixed effects. The second is an intensity model for triggered draws as a function of recent counts, clinical events, and covariates. The third is a survival model for death. The central correction weights each observed count by the inverse of its estimated probability of being observed:

$$w_{ij} = \frac{1}{\hat\lambda_i(t_{ij} \mid H_i(t_{ij}^-))},$$

Here $\hat\lambda_i$ is the fitted rate at which patient $i$ has a triggered draw at time $t_{ij}$, and $H_i(t_{ij}^-)$ is that patient's recorded history just before the draw. Scheduled draws receive weights from the separate, far flatter scheduled-visit process. Informative dropout from surveillance is handled with standard inverse-probability-of-censoring weights. Confounding by chemotherapy regimen, which the breast cancer work shows strongly affects lymphopenia risk (Potential Determinants for Radiation-Induced Lymphopenia in Patients With Breast Cancer Using Interpretable Machine Learning Approach), is addressed with propensity weighting over regimen when estimating the effect of spleen dose. This follows the balancing strategy used for treatment comparisons in Long-Term Survival in Men With High Grade Prostate Cancer: A Comparison Between Conservative Treatment, Radiation Therapy and Radical Prostatectomy—A Propensity Scoring Approach.

We will report three things. The first is corrected estimates of the fraction of patients still lymphopenic at 6, 12, 24, and 36 months. The second is the dose-response of corrected 12- and 24-month counts on spleen $V_5$, including whether the 180.6 and 272.2 cm³ cutoffs remain the relevant region. The third is, secondarily for the lung cohort, whether lymphocyte levels after progression predict survival after progression. That analysis is a direct test of one candidate explanation for why counts predicted overall but not progression-free survival in the Sun et al. cohort.

Baselines

The do-nothing reference is an unweighted mixed-effects trajectory fitted to all available counts. This is what a careful analyst would produce by simply extending follow-up. It uses the same data and covariates as our model and differs only in the weights, so any difference between the two is attributable to the testing correction.

The strongest existing method for this specific question is the day-120 landmark approach of Ma et al., which uses the recovery index and spleen $V_5$ cutoffs in a Cox model. The reported figures to beat or revise are 87.4% persistent lymphopenia at day 120 and a 60.9% reduction in risk of death for $V_5 < 272.2$ cm³ (Clinical Considerations of Splenic Dose Constraints to Mitigate Radiation‐Induced Lymphopenia). No reported discrimination metric such as a concordance index for survival prediction in a lymphopenia cohort was found on Lacuna. We will refit this landmark model on the extended cohort, with its original variables, as the reference. "Beating" it here means two things. On the lung validation window, our method must estimate long-term recovery with lower error. On survival prediction, corrected trajectories must improve time-dependent concordance and Brier score over the landmark model at 2- and 3-year horizons. We will tune both on the same folds.

Three further competitors receive the same covariates and cross-validation folds. The first is multiple imputation of missing counts under a missing-at-random assumption, followed by the same trajectory model. The second is a time-dependent Cox model using the last observed count. The third is CoxSig+, whose authors report a concordance index of 0.867 against 0.813 for Dynamic DeepHit on NASA turbofan maintenance data (Dynamic Survival Analysis with Controlled Latent States). No reported result was found on Lacuna for CoxSig on radiotherapy lymphocyte data. For reference only, XGBoost reached an ROC-AUC of 0.841 for acute lymphopenia in breast cancer (Potential Determinants for Radiation-Induced Lymphopenia in Patients With Breast Cancer Using Interpretable Machine Learning Approach). That is a different, acute endpoint, and we will not claim to beat it.

The ablation that isolates our key ingredient fits a single intensity model to all draws without the scheduled/triggered label. If this ablation performs as well as the full model, the labeling effort is not justified.

Why not just…?

Why not just keep the day-120 recovery index? It is simple and already predicts five-year survival (Clinical Considerations of Splenic Dose Constraints to Mitigate Radiation‐Induced Lymphopenia). But with 87.4% of patients still lymphopenic at that point, the index cannot separate slow recoverers from those whose lymphocyte counts never recover, and a spleen constraint aimed at long-term immune capacity needs that distinction. If corrected 24-month counts are nearly fully explained by the day-120 index, this proposal's clinical payoff shrinks to confirming the landmark. We will report that result plainly.

Why not just multiple imputation or a standard joint model of trajectory and survival? Both assume that, given the recorded history, whether a count exists tells us nothing about its value. Oncology testing violates this (Risk factors for steroid-refractory in immune checkpoint inhibitor-induced colitis: a retrospective cohort study). The lung infusion-cycle window lets us measure how far each method lands from the near-complete trajectory, instead of arguing the point.

Why not just restrict the analysis to scheduled surveillance draws? This is the most serious threat, because it is cheap and removes most triggered-draw bias. Its weakness is that scheduled surveillance itself thins out informatively: patients who are doing well get discharged to primary care, and some very sick patients stop attending. It also discards data. We include it as an explicit baseline. If it matches the full model on the validation window, we will recommend it, and the contribution becomes the labeled long-term data and the revised dose-response rather than the weighting machinery.

Why not a continuous-time model such as CoxSig or a deep state-space model? These models handle irregular timing well. However, they learn from how often a patient is tested, which improves prediction while contaminating any estimate of how dose affects recovery. With roughly 150 patients per cohort, their capacity is also a liability. We run CoxSig+ as a competitor for prediction. For estimating recovery and dose effects, we expect the corrected mixed model to be preferable.

Why not a prospective spleen-sparing trial? A trial is the right end point, but it needs a target. Choosing between a $V_5$ limit calibrated at 120 days and one calibrated at two years is exactly what this study decides, and the study can do so with existing records.

Proposed Experiments

The decisive check uses the lung consolidation window. We will hide the protocol-driven pre-infusion counts, keep only the triggered and off-cycle draws, and ask each method to reconstruct the population mean trajectory and the fraction of patients below 1.0 × 10⁹/L at each cycle. Error against the hidden counts is the primary metric. Our method succeeds if it reduces this error relative to the unweighted and imputation baselines by more than the bootstrap interval width. A second check is a negative control: the correction should leave estimates essentially unchanged for platelet counts in the lung window, where testing bias is weaker. The third, substantive analysis compares corrected and naive long-term persistent-lymphopenia rates in the gastric cohort, and tests whether spleen $V_5$ still predicts corrected 24-month counts after propensity weighting over chemotherapy regimen. A material shift would mean current constraints are calibrated to the wrong horizon, for example if the naive curve overstates persistence by more than ten percentage points, or if the dose-response flattens after 12 months. Power is the main risk. Long-term counts in a single-center cohort of about 150 patients will be sparse. We will therefore report confidence intervals for every horizon and treat 36-month estimates as exploratory unless at least half of surviving patients have one or more scheduled draws after 24 months.