References for the immune biomarkers and infection risk project

What to read, in what order, and what has actually been checked

The reading queue for a project on longitudinal immune biomarkers and infection risk: the papers that quantify what a neutrophil, lymphocyte or IgG level does to infection risk, the model-based predictions already published, and a status marker on every entry.

Nothing here has been read. The queue was assembled on 2026-09-09 from literature searches, so every entry is a characterization of a paper rather than a reading of one, and every entry carries ❌. Bibliographic details (authors, journal, year, volume, pages, DOI) were fetched from Europe PMC on the same day and are correct. What each paper says was taken from its abstract or from a search summary, and the numbers quoted below are quoted for the purpose of deciding what to read, not for use in any analysis.

The working document is the specification, and the folder’s other documents are on the project index.

Status markers

• ❌ Not checked. No claim in the project has been verified against this source. Where the project characterizes it, the characterization comes from an abstract or a search result.
• ⚠️ Transcribed, unverified. A number, table or model structure in the project came from this source but has not been checked against it.
• ✅ Checked. Verified against the source.

On a reading-queue entry, ✅ means the question posed against the entry has been answered, by reading the paper or by deciding it is out of scope, and the note under the entry says which. On a source entry below the queues, ✅ keeps the stricter meaning: the claim the project draws from it has been checked against the source. None is ✅ yet.

How to read these

Read against the question under each entry rather than through the paper front to back. The general method is on the Reading Papers page: three passes, stopping at the pass that answers the question.

Where to start

Five papers decide whether to start the project, in this order.

1. Entry 2, Li 2016. The one paper found that pools several studies and reports infection risk per day of severe neutropenia. It is the shape of the answer this project wants, on one biomarker.
2. Entry 11, Netterberg 2018. A turnover model for a biomarker time course feeding a parametric time-to-event model for febrile neutropenia. The method the project would use, already published on adjacent biomarkers.
3. Entry 4, Derfuss 2024. Six thousand patients pooled across thirteen ocrelizumab trials, with IgG below the lower limit of normal related to serious infection rate. The closest published thing to the proposed analysis, and it is a categorized level rather than a trajectory.
4. Entry 12, Rejeski 2021. A baseline score, not a trajectory, that predicts severe infection after CAR-T. Any longitudinal model has to beat a score this simple.
5. Entry 16, Rizopoulos 2017. Joint modelling against landmarking. The fork the project has to pick before it fits anything.

Reading queue: what a biomarker level does to infection risk

These establish the association the project would model, and each reports it in a different form: a threshold, a duration, a categorized level, or a randomized intervention on the biomarker itself.

1. Bodey et al. 1966 — circulating leukocytes and infection in acute leukemia ❌

Annals of Internal Medicine 64(2):328–340. doi:10.7326/0003-4819-64-2-328 · PMID 5216294

The origin of the neutrophil thresholds still in use. Reported as establishing that infection incidence rises as the count falls below 1000, 500 and 100 cells/mm³, and as reporting a rate of decline and a duration effect alongside the level. Read it for the shape of the original data: whether the level, the slope and the duration were separable in 1966, and on how many patients.

2. Li et al. 2016 — severity and duration of chemotherapy-induced neutropenia and infection ❌

Supportive Care in Cancer 24(10):4377–4383. doi:10.1007/s00520-016-3277-0 · PMID 27278272

Pools 271 patients from six clinical trials with serial neutrophil measurements, and reports a hazard of infection-related hospitalization rising by roughly 28 to 30% per additional day of severe neutropenia. Pooling across studies to relate a biomarker time course to an infection event is what this project proposes, and this is it, done on one biomarker in one disease setting. Read it first, and read it for the mechanics: how the six trials were harmonized, how duration was derived from scheduled counts, and what the model of the hazard was.

3. Warny et al. 2018 — lymphopenia and infection in 98,344 individuals ❌

PLoS Medicine 15(11):e1002685. doi:10.1371/journal.pmed.1002685 · PMID 30383787

A Danish population cohort relating blood lymphocyte count to hospitalization for infection and to infection-related death, reported as showing a graded relationship down to counts below 0.5 x 10⁹/L. Untreated general population, so it separates the biomarker from the drug that would otherwise have depleted it. Read it for the confounding structure, which is the question the pooled analysis will face in a different form.

4. Derfuss et al. 2024 — infections and risk factors across 13 ocrelizumab trials ❌

Therapeutic Advances in Neurological Disorders 17:17562864241277736. doi:10.1177/17562864241277736 · PMID 39399100

A pooled analysis of 6,155 patients, reported as relating IgG below the lower limit of normal to serious infection rates, alongside disability score, comorbidity and body weight. Read it for what a sponsor can do with pooled individual patient data from its own programme, and for what it stopped short of: the analysis appears to categorize IgG rather than model its trajectory, which is the gap this project would work in.

5. Saidha et al. 2023 — immunoglobulin trends under anti-CD20 antibodies in multiple sclerosis ❌

Neurological Sciences 44(5):1515–1532. doi:10.1007/s10072-022-06582-y · PMID 36648561

A systematic review of how IgG, IgA and IgM move over time under rituximab, ocrelizumab and ofatumumab. The trajectory half of the problem without the event half. Read it for the inventory of studies reporting serial immunoglobulin measurements, which is a candidate source list for the pooled dataset.

6. Huang et al. 2026 — infection risk with teclistamab, systematic review and meta-analysis ❌

Frontiers in Immunology 17:1804838. doi:10.3389/fimmu.2026.1804838 · PMID 42148087

Five studies and 714 patients on a BCMA-directed T-cell engager, reported as finding any-grade infection in 76.4% of trial patients against 45.4% in real-world cohorts, with heterogeneity attributed to follow-up duration and to how much immunoglobulin replacement was given. Read it for the definition problem: whether the infection endpoints across those five sources are the same endpoint. If they are not, the same obstacle applies to any pooled analysis.

7. Gale et al. 1988 — intravenous immunoglobulin for prevention of infection in CLL ❌

New England Journal of Medicine 319(14):902–907. doi:10.1056/NEJM198810063191403 · PMID 2901668

The randomized trial that raised IgG on purpose and counted infections. Every observational entry above is confounded by the fact that a sicker patient has both a lower IgG and a higher infection risk; this one is not. Read it to learn what size of effect an intervention on the biomarker produced, since that is the ceiling on what a predictive model could deliver.

8. Carrillo de Albornoz et al. 2025 — immunoglobulin use, survival and infection in CLL ❌

Blood Advances 9(20):5367–5377. doi:10.1182/bloodadvances.2025015867 · PMID 40742276

A linked-hospital-data cohort of 6,217 patients, reported as finding a higher infection incidence during immunoglobulin replacement periods than outside them, and as concluding that the causal relationship needs further work. Confounding by indication, stated plainly in a recent paper. Read it against entry 7 and note what the two designs do to the same question.

Reading queue: model-based prediction

9. Friberg et al. 2002 — myelosuppression model with parameter consistency across drugs ❌

Journal of Clinical Oncology 20(24):4713–4721. doi:10.1200/JCO.2002.02.140 · PMID 12488418

Proliferating compartment, three transit compartments, circulating compartment, feedback. The trajectory submodel for neutrophils is a solved problem and this is the solution, with system parameters that held across several drugs. Read it for whether the same structure has any claim on B cells or immunoglobulins, whose turnover is slower and whose recovery after depletion is not obviously a feedback loop of the same kind.

10. Netterberg et al. 2017 — model-based prediction of myelosuppression from frequent monitoring ❌

Cancer Chemotherapy and Pharmacology 80(2):343–353. doi:10.1007/s00280-017-3366-x · PMID 28656382

Forecasts an individual patient’s neutrophil time course from their own measurements so far, reported as reaching sensitivity of 90% or better for grade 4 neutropenia before it occurs, with accuracy worst around the nadir. This is the individual-prediction half of the project, on the biomarker where the model already exists. Read it for how far ahead a forecast stays useful, which sets how often the biomarkers would have to be drawn.

11. Netterberg et al. 2018 — IL-6 and CRP time courses predicting febrile neutropenia ❌

British Journal of Clinical Pharmacology 84(3):490–500. doi:10.1111/bcp.13477 · PMID 29178353

Turnover models for two biomarker time courses, linked to parametric time-to-event models for febrile neutropenia, with the IL-6 time course the better predictor and baseline age much weaker. The exact architecture this project proposes, published, on inflammation markers rather than on immune cell counts. Read it second, and read it for the link function: which feature of the time course entered the hazard, and what was tried and rejected.

12. Rejeski et al. 2021 — CAR-HEMATOTOX ❌

Blood 138(24):2499–2513. doi:10.1182/blood.2020010543 · PMID 34166502

A score built from pre-treatment haematopoietic reserve and inflammation, predicting prolonged neutropenia and severe infection after CD19 CAR-T. Baseline values only, no trajectory, and it is in clinical use. This is the comparator: a longitudinal model earns its complexity only by beating a scoring sheet computed once, before treatment.

13. de Boer et al. 2025 — population-based validation of CAR-HEMATOTOX ❌

Blood Advances 9(21):5641–5650. doi:10.1182/bloodadvances.2025016689 · PMID 40668622

External validation of entry 12 for hematotoxicity, infections and survival in a national cohort. Read it for the discrimination the score actually achieves outside its development set, which is the number a new model would be measured against.

14. Sheehy et al. 2025 — clinical prediction models for febrile neutropenia, systematic review ❌

Supportive Care in Cancer 33(7):537. doi:10.1007/s00520-025-09562-y · PMID 40467900

An inventory of the prediction models already published for one infection outcome. Read it to find out how many of them use serial measurements rather than baseline covariates, and how they were validated. If the answer is that almost none use trajectories, that is the argument for the project. If several do and none beat a baseline model, that is the argument against.

Methods for linking a trajectory to an event

15. Rizopoulos 2011 — dynamic predictions and prospective accuracy in joint models ❌

Biometrics 67(3):819–829. doi:10.1111/j.1541-0420.2010.01546.x · PMID 21306352

Updating a patient’s predicted risk as each new measurement arrives, and measuring whether those updates are accurate. The JMbayes2 R package implements this family. Read it for the accuracy measures, since a time-updated prediction cannot be scored with a single AUC.

16. Rizopoulos, Molenberghs and Lesaffre 2017 — joint modelling against landmarking ❌

Biometrical Journal 59(6):1261–1276. doi:10.1002/bimj.201600238 · PMID 28792080

The two ways to put a time-varying covariate into a survival model, compared directly. Landmarking is simpler and pools across studies more easily; joint modelling handles measurement error and informative visit timing. Read it before committing, because the choice decides what the pooled dataset has to look like.

Data sources

Marked ❌ throughout: none has been queried, and what each holds for longitudinal immune laboratory values is unknown.

• Vivli (https://vivli.org). Individual participant data from completed trials, analyzed inside a secure environment rather than downloaded. Reported as holding over 7,500 trials. The question to put to it: how many studies expose serial haematology and immunoglobulin panels together with coded infection adverse events.
• Project Data Sphere (https://projectdatasphere.org). Patient-level data from oncology trials, historically phase 3 comparator arms.
• The YODA Project (https://yoda.yale.edu). A third route to participant-level data, with a different sponsor list.
• Published figures. Where individual data cannot be obtained, mean trajectories digitized from publications support a model-based meta-analysis of the biomarker time course, but not a within-patient link to an event.

Sources for the numbers in the project

• ❌ The thresholds. The index and Section 4 of the specification name an absolute neutrophil count below 0.5 x 10⁹/L and an IgG in the 400 to 500 mg/dL range as levels at which practice already acts. The neutrophil figure traces to entry 1 by way of common usage. The immunoglobulin figure came from search summaries citing NCCN guidance for CLL and general secondary-immunodeficiency practice, and no guideline text has been read. Both need a source before either is used as a comparator.

What was searched and did not turn up

Searched on 2026-09-09: meta-analyses of immunoglobulin level against infection, neutropenia severity and duration against infection, pharmacometric and time-to-event models of febrile neutropenia, joint models of a longitudinal biomarker with a time-to-infection outcome, infection prediction after CAR-T and after bispecific antibodies, and individual patient data sharing platforms.

Nothing found pools several studies, across drug classes, into one model where a trajectory of an immune biomarker enters the hazard for infection. The pieces exist separately: entry 2 pools studies for one biomarker, entry 11 links a trajectory to an event in one trial, entry 4 pools a programme but categorizes the biomarker. Whether the absence is real or an artifact of searching in English on abstracts is the first thing to establish, and a search of the pharmacometrics meeting abstracts (PAGE, ACoP) has not been run.

Also not searched yet: paediatric immunology and primary immunodeficiency, where immunoglobulin trajectories are measured for decades; transplant immune reconstitution, where CD4 recovery curves against infection are an established literature; and HIV, where the CD4 count against opportunistic infection is the oldest worked example of this exact problem.