Bridging gaps across the CLL care pathway – Part 2: Knowing When To Treat

November 20, 2025

Clinical Development - Commercial Excellence - Diseases - Insights - Medical Affairs - Oncology - Predicting patient outcomes - Value & Access HEOR

By Léon Van Wouwe, Clinical Innovation Director, Volv Global

Knowing When To Treat – AI for Transformation, Infection Risk, and Adaptive CLL Care

Introduction: the long arc of CLL and its inflection risks

Most people who have been diagnosed with chronic lymphocytic leukaemia (CLL) have no symptoms initially and are usually advised that no treatment is necessary upon diagnosis. Since the disease is a chronic, slow-growing cancer, many have no symptoms and experience little to no change in their health for many years. So, actively monitoring the disease without starting treatment has been proven to be a safe standard of care. [1]

However, their lived clinical experience is rarely straightforward. Key inflection points – transformation (to aggressive lymphoma), high infection risk, or progression to second-line therapy – define outcomes, costs, and quality of life. The cumulative lifetime cost of CLL in the U.S. is estimated at ~USD 1,142,357 (for a 70-year-old entry cohort), with ~8.03 years modelled time in CLL care, and most of that time (6.0 years) in first-line therapy. [2] Targeted therapies drive ~77 % of total cost; inpatient/ED care ~20 %. Real-world U.S. data show that treated patients incur ~USD 101,122 in annual costs. [3]

While it is best to wait to begin treatment until symptoms are present, it is important not to wait until the symptoms are severe. Given these stakes, improved prediction of transformation and infection, and the timing of when to initiate second-line therapy, are pressing unmet needs – and this is a care-gap that AI-enabled longitudinal models are uniquely poised to bridge.

The status quo: diagnostic inertia and inefficiencies

Today, most CLL diagnoses arise from incidental lymphocytosis on a CBC or differential, followed by haematology referral and confirmatory flow cytometry (≥ 5 ×10⁹/L clonal B cells, sustained) with immunophenotyping. [1] Patients who present with nonspecific symptoms (fatigue, night sweats, low-grade fevers, recurring infections, lymphadenopathy) may traverse multiple outpatient encounters before evaluation. Because lymphocytosis has many benign causes, clinicians may not act until trends are evident.

This reactive workflow causes delays (weeks to months) and uneven referral patterns, especially when absolute lymphocyte count (ALC) is only modestly elevated or fluctuating. In non-academic settings, molecular and cytogenetic testing may not be readily available or may have long turnaround times, adding friction to early decision-making.

Richter transformation: the black swan event crucial to recognise

Richter’s transformation (RT), a rare, aggressive histologic conversion of CLL into a high-grade lymphoma – typically to diffuse large B-cell lymphoma (~90% of RT cases) – is a devastating event with high mortality. [4] RT may occur any time, with median intervals between 1.8 and 5 years in reported series.

Clinically, RT may present with sudden lymph node growth, B symptoms, elevated LDH, or imaging lesions (PET-avid). Yet biopsy and PET/CT are costly, not always timely, and typically done after symptom onset. Detecting RT risk earlier is a major challenge.

Algorithmic models can help by integrating longitudinal laboratory test result trends, clinical symptoms, imaging signals (if available), and perhaps MRD/NGS readouts to generate a “transformation risk score.” That score could flag patients for early PET or biopsy, enabling earlier intervention or trial enrolment.

With an overall prevalence of RT of about 2-10% in CLL patients according to multiple published studies, RT is a rare black swan event – an unpredictable occurrence that has severe consequences, often rationalised only in hindsight as if it could have been expected. [4]

Therefore, predictive models to warn for it must handle class imbalance, avoid overfitting, and maintain robust interpretability. But even moderately predictive scores could shift a few high-risk cases earlier, with outsized incremental value.

Infection risk: an ever-present shadow

Infections remain a leading driver of morbidity and mortality in CLL, both from disease-associated immunodeficiency and therapy-related immunosuppression. [5]

Some recently published models in CLL have sought to stratify risk of ≥ grade 3 infection. For example, one XGBoost model in a cohort (~n = 647) distinguished high- vs low-risk patients (57% vs 28% risk) with a c-index of ~0.687. [6]

Integrating infection risk scores into care offers several advantages:

Prophylactic decisions (antimicrobials, immunoglobulin support).
Adjusted therapy dosing or scheduling.
Enhanced monitoring or early-warning alerts in high-risk windows.

Second-line initiation and dynamic adaptation

Transition to second-line therapy is usually reactive – triggered by progression or intolerance. But in a future care model, one might envisage predictive TTNT2 (time-to-next-treatment) curves adapted per patient. Longitudinal data (response kinetics, residual disease, lab trends, comorbidities) could feed models that forecast when a patient is likely to require escalation, allowing pre-emptive therapy planning or trial enrolment.

Unsupervised ML methods or survival-optimised cluster models may identify latent subgroups with distinct progression kinetics, beyond standard prognostic indices. For instance, in one study, unsupervised clustering of immunophenotype/genetic data in early-stage patients uncovered risk trajectories not obvious under classical staging. [7]

A continuous-risk “AI companion” for CLL

The vision is of a real-time, evolving CLL risk engine:

Continuously updating risk probabilities for RT, infection, progression to next line.
Using explainable machine learning (Shapley values, feature attribution) to surface driving signals.
Escalating actionable flags to clinicians (e.g., “Consider PET”, “Prophylaxis recommended”, “Prepare for 2L enrolment”).
Periodic retraining using federated data across systems to adapt to shifting drug landscape.

From a pharma standpoint, this enables trial-ready risk cohorts, dynamic enrichment, and more efficient post-launch monitoring (e.g. real-world risk-adjusted benchmarks).

Challenges and guardrails

Prospective validation is essential – retrospective models alone will not persuade.
Explainability and clinician trust are critical: opaque “black box” models risk rejection.
Data heterogeneity and drift: models must adapt as patterns shift (new therapies, demographics).
Low event rates (RT) impose statistical limitations – even the best models may have limited positive predictive value.

Conclusion: toward anticipatory haematology

In CLL, the greatest value may lie not in retrospective prognostication but in anticipation. By looking at AI as a dynamic companion – that continually monitors, predicts, and prompts –we can reframe CLL care for HCPs from reactive to pre-emptive, making better use of therapy windows so that for patients, we are optimising outcomes and reducing crises. For pharma, the payoff is smarter trials, timely enrolment, and evidence-based risk-adjusted deployment of therapies.

About the author

Léon van Wouwe has 20+ years’ global experience in clinical development and operations, uniting data science with pharma and research. He drives cross-functional collaboration to advance innovative treatments.