📄 medicine

Toward uncertainty-aware clinical decision support for treatment response prediction in metastatic NSCLC: integrating FDG-PET, T-cell repertoire, and cytokines with conformal prediction

This study presents a multimodal clinical decision support framework that integrates longitudinal FDG-PET, T-cell receptor repertoire, and cytokine biomarkers with conformal prediction to generate uncertainty-quantified early treatment response predictions for metastatic NSCLC, demonstrating significantly superior accuracy and reliability compared to the current PD-L1 standard.

Original authors: Faisal Yaseen, Daniel S. Hippe, Sunan Cui, Jie Fu, Yejin Kim, Clemens Grassberger, Lei Deng, Ting Ye, Paul E Kinahan, Jing Zeng, John H Gennari, Stephen R. Bowen

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Faisal Yaseen, Daniel S. Hippe, Sunan Cui, Jie Fu, Yejin Kim, Clemens Grassberger, Lei Deng, Ting Ye, Paul E Kinahan, Jing Zeng, John H Gennari, Stephen R. Bowen

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a patient is diagnosed with metastatic non-small cell lung cancer, the path forward is often a difficult gamble. Doctors typically prescribe a combination of chemotherapy and immunotherapy, a treatment that has saved many lives but fails to help nearly half of those who receive it. The current standard for deciding who gets this treatment relies on a single biological marker called PD-L1, a protein found on tumor cells. However, this marker is an imperfect guide; it often fails to distinguish clearly between patients who will respond well and those who will not, leaving many to endure toxic side effects without any clinical benefit. The medical community has long sought a way to predict the outcome of these therapies earlier and more accurately, hoping to spare patients from ineffective treatments while identifying those who truly need them. To do this, researchers have begun looking beyond a single snapshot of the disease, instead gathering a wide array of signals from the body, including how the tumor consumes energy, how the immune system is organized, and what chemical messages it is sending.

A team of researchers at the University of Washington and the Fred Hutchinson Cancer Center has taken a significant step toward solving this problem by building a new kind of prediction tool. They focused on thirty-five patients with advanced lung cancer who were about to begin their first round of chemoimmunotherapy. Instead of relying on just one type of data, the team collected three distinct streams of information from each patient at two different times: before treatment started and after the first cycle was complete. The first stream came from a specialized scan called an FDG-PET, which shows how actively the tumor is eating sugar, a sign of its metabolic energy. The second stream involved analyzing the blood for the diversity of T-cell receptors, which act as the immune system's unique identification tags, revealing how varied and ready the body's defenses are. The third stream measured a panel of cytokines, which are small proteins that act as messengers to coordinate inflammation and immune activity. By weaving these three different types of data together, the researchers aimed to create a more complete picture of what was happening inside the patient.

The researchers faced a critical challenge: medical predictions must not only be accurate but also honest about their own uncertainty. In the past, computer models often forced a binary answer, declaring a patient a "responder" or a "non-responder" even when the evidence was weak or conflicting. This study introduced a method called conformal prediction, which allows the model to say "I am not sure" when the data is ambiguous. Rather than guessing, the system can flag a patient for closer monitoring or further testing if the signals from the scan, the immune cells, and the blood markers do not agree. This approach ensures that when the system does make a prediction, the medical team can trust the confidence level behind it. The team tested their framework against the current clinical standard, the PD-L1 score, and found that the new multimodal approach was far superior. While the PD-L1 score performed no better than a random guess in this group, the new model, which combined the PET scan with the immune data, correctly distinguished between responders and non-responders with much higher accuracy.

The study revealed that the most useful information changes over time. Before treatment began, the diversity of the T-cell receptors in the blood was the strongest single predictor of who would respond. However, once the patients had undergone one cycle of treatment, the story shifted. At that mid-point, the levels of specific inflammatory proteins in the blood became the most powerful indicator, while the T-cell data became less distinct. By combining the PET scan with these changing blood markers, the researchers achieved their best results. The system successfully predicted outcomes with an accuracy that significantly exceeded the current standard of care. Perhaps most importantly, the model was able to identify patients for whom a confident prediction could not be made. In these cases, the system did not force a wrong answer but instead signaled that the evidence was insufficient, effectively protecting the patient from a potentially harmful misclassification.

To demonstrate that this complex mathematical framework could be used in a real hospital setting, the team built a prototype interface. This digital tool takes the patient's scan and blood test results and displays a clear prediction along with a measure of how certain the system is. In some cases, the tool resolved confusion that existed when looking at a single test in isolation, providing a definitive answer where previous methods had been unsure. In other cases, it correctly identified when conflicting signals meant that a decision should be delayed. The researchers acknowledge that their study involved a relatively small group of patients from a single location, meaning these findings are a promising proof of concept rather than a final solution. Independent validation with larger groups will be necessary to confirm that these results hold true for everyone. Nevertheless, this work offers a tangible vision of the future of cancer care, where treatment decisions are guided by a comprehensive, uncertainty-aware view of the patient's biology, ensuring that every prediction is as reliable and honest as the data allows.

Technical Summary: Toward Uncertainty-Aware Clinical Decision Support for Treatment Response Prediction in Metastatic NSCLC

Problem Statement
Patients with metastatic non-small cell lung cancer (mNSCLC) exhibit highly variable responses to first-line chemoimmunotherapy, with nearly half deriving limited clinical benefit despite exposure to toxicity and financial burden. Current clinical standards, primarily the PD-L1 tumor proportion score (TPS), demonstrate limited discriminative accuracy in predicting treatment response. Furthermore, single-modality biomarkers fail to capture the complex interplay between tumor burden, immune response, and systemic inflammation. A critical gap exists in translating predictive models into clinical decision support systems (CDSS): models often lack calibrated probability estimates and fail to quantify uncertainty, forcing binary decisions even when evidence is ambiguous. This study addresses the need for a reliable, uncertainty-aware early response prediction framework that integrates multimodal biomarkers and explicitly communicates prediction confidence to clinicians.

Methodology
The study utilized data from 35 patients with mNSCLC enrolled in the prospective PET-BRIGHT trial (NCT04151940), who received first-line carboplatin–pemetrexed–pembrolizumab or carboplatin–paclitaxel–pembrolizumab. Biomarkers were collected at two timepoints: baseline (PreTx) and week 3 after one cycle of therapy (MidTx).

Multimodal Data Integration: The framework integrated three distinct data modalities:
1. FDG-PET/CT: Metabolic tumor volume (MTV) metrics, specifically standardized uptake values (SUVmax, SUVpeak) and total lesion glycolysis (TLG).
2. T-cell Receptor (TCR) Repertoire: Diversity and clonality metrics (e.g., Pielou evenness, Simpson clonality) derived from peripheral blood mononuclear cells.
3. Cytokines: Levels of 14 circulating inflammatory and immunomodulatory cytokines.
Model Development:
- Feature Selection: A nested leave-one-out cross-validation (LOOCV) approach was employed. An inner loop selected the single best-performing feature per modality based on AUROC, which was then used in the outer loop to fit the model. This maintained a favorable events-per-predictor ratio.
- Classification: Class-balanced logistic regression was used to generate probability estimates.
- Fusion Strategies: Both early fusion (concatenating features) and late fusion (averaging modality-specific probabilities) were evaluated across unimodal, bimodal, and trimodal combinations.
Uncertainty Quantification: Conformal prediction was applied to convert model probabilities into prediction sets with finite-sample validity guarantees. Targeting an 80% reliability level ( $\alpha = 0.20$ ), the system outputs either a single class (decisive prediction) or both classes (ambiguous prediction) for each patient. This allows the system to abstain from making a definitive call when evidence is insufficient.
Statistical Evaluation: Performance was assessed via AUROC and Brier score (calibration). Statistical significance was determined using permutation testing against an empirical null distribution (500 permutations) to account for feature selection optimism. The models were benchmarked against the clinical standard of PD-L1 TPS dichotomized at 1%.

Key Results

Benchmark Performance: The current clinical standard, PD-L1 TPS, discriminated poorly between responders and non-responders (AUROC 0.58, 95% CI 0.39–0.78), performing no better than chance in this cohort.
Unimodal Performance:
- PreTx: TCR repertoire was the strongest single modality (AUROC 0.80), followed by FDG-PET (0.76) and cytokines (0.54).
- PreTx + MidTx: The predictive landscape shifted; cytokines became the strongest single modality (AUROC 0.78), driven by week-3 IL-6 levels, while TCR performance attenuated (0.66). FDG-PET remained stable at 0.76.
Multimodal Performance:
- Late fusion consistently outperformed early fusion.
- The highest PreTx discrimination was achieved by late fusion of PET and TCR (AUROC 0.85).
- The highest overall discrimination (PreTx + MidTx) was achieved by late fusion of PET and cytokines (AUROC 0.86).
- Thirteen of 22 model/timepoint combinations significantly exceeded the permutation null ( $p < 0.05$ ).
Uncertainty and Calibration:
- Multimodal fusion improved calibration (lower Brier scores) and decisiveness compared to unimodal models.
- The singleton rate (proportion of patients receiving a single-class prediction) increased from 78% (unimodal TCR at PreTx) to 89% (PET+TCR late fusion at PreTx).
- Empirical coverage rates were 78% and 82% against the 80% target, validating the reliability of the conformal prediction layer.
Clinical Utility Examples: Representative cases demonstrated how multimodal integration could resolve ambiguity (e.g., clarifying a PET-only ambiguous case) or correct baseline misclassifications using mid-treatment cytokine data. Conversely, the framework successfully flagged cases with conflicting signals as uncertain, preventing incorrect definitive predictions.

Significance and Claims
The authors claim that this study represents one of the first prospective efforts to integrate longitudinal FDG-PET imaging with circulating immune biomarkers (TCR and cytokines) under a conformal prediction framework for mNSCLC. The primary significance lies in the development of a multimodal framework that:

Outperforms the current clinical standard (PD-L1) in discrimination and calibration, with all multimodal models exceeding the PD-L1 benchmark by margins ranging from 0.12 to 0.28 AUROC.
Quantifies patient-level uncertainty, explicitly identifying patients for whom a confident prediction cannot be made, thereby supporting safer clinical reasoning.
Demonstrates feasibility for clinical translation through a prototype browser-based Clinical Decision Support (CDS) interface that visualizes probabilities and prediction sets.

The paper maintains a modest stance regarding its findings, noting that while multimodal models exceeded the PD-L1 benchmark, the incremental benefit of multimodal fusion over the best single modality could not be established statistically due to the limited sample size ( $N=35$ ). Consequently, the reported performance gains of fusion over the strongest single modality are presented as suggestive rather than definitive, and the study emphasizes the need for independent validation and prospective evaluation to confirm generalizability and clinical impact. The work highlights that while multimodal models improve accuracy, the true clinical value is enhanced by the ability to signal when a model should not be trusted, a feature absent in conventional binary classifiers.

Technical Summary: Toward Uncertainty-Aware Clinical Decision Support for Treatment Response Prediction in Metastatic NSCLC

More like this