Immune biomarkers and infection risk

Longitudinal models of neutrophils, B cells and immunoglobulins

Placeholder for a project on pooling immune biomarker time courses across studies to model infection risk. Nothing is written yet; this page states the question and what has to be settled before the work starts.

Nothing has been fitted. The question and the analysis that would answer it are specified; no dataset has been obtained and no model has been run.

The question. Therapies that deplete neutrophils, B cells or immunoglobulins raise the risk of infection, and the laboratory values that track the depletion are drawn at nearly every visit. Absolute neutrophil count (ANC), CD19+ B cells, and immunoglobulin G, A and M (IgG, IgA, IgM) are the usual ones. Can a longitudinal model of those time courses predict which patient is heading for a serious infection, precisely enough to change when a patient is monitored, when prophylaxis starts, or whether the next dose is given?

What would be modelled. A trajectory model for each biomarker as a function of dose, schedule and time, linked to a time-to-event model for infection. The question the link answers is which feature of the trajectory carries the risk: the current value, the rate of change, or the cumulative time spent under a threshold. Thresholds such as an absolute neutrophil count below 0.5 x 10⁹/L, or an IgG in the 400 to 500 mg/dL range at which immunoglobulin replacement is considered, are already used in practice as if the answer were the current value, and they are the comparator any model has to beat.

Why pooling across studies. Grade 3 or higher infections are uncommon enough that a single early-phase cohort produces a handful of events, which will not support a model with several biomarkers in it. The events exist across trials, drug classes and sponsors: anti-CD20 antibodies, T-cell engagers, CAR-T, and cytotoxic chemotherapy all deplete some part of the same system. Whether those studies can be pooled into one model, or only into one meta-analysis of separate fits, is the first thing the project has to answer.

Out of scope. This is not a mechanistic model of immune reconstitution, not a vaccine-response analysis, and not a dosing recommendation for any individual drug.

What has to be settled before starting

1. The data. Which pooled sources are reachable, and whether they carry individual patient records or only published summary curves. Vivli, Project Data Sphere, the YODA Project and figures digitized from publications are the candidates, and they differ in what a longitudinal model can be fit to.
2. The infection endpoint. Whether infection is recorded consistently enough across sources to be one endpoint. Grading, adjudication, and the distinction between a documented pathogen and a febrile episode all vary.
3. Assay comparability. Whether a B cell count from flow cytometry at one site is the same measurement as at another, and what the lower limit of quantification does to trajectories that go to zero and stay there.
4. Whether it has been done. The searched literature has each piece separately: one pooled analysis relating days of severe neutropenia to infection risk, one trial-level model linking a biomarker time course to febrile neutropenia, and one sponsor programme pooling immunoglobulin levels across thirteen studies. None combines them. References says what each leaves open.

Documents

• Working specification. The population, the map of what links to infection, the biomarkers, the infection endpoint, the model and its comparator, the confounding, and the sequence to build it in. Its In brief section is one screen and says what the rest argues.
• References. Sixteen papers in three queues: what a neutrophil, lymphocyte or immunoglobulin level does to infection risk, the model-based predictions already published, and the two ways to link a trajectory to an event. Nothing has been read, so every entry carries ❌, and the top of the page says where to start.