🔬 oncology

Forecasting patient-specific tumor response using patient-reported outcomes in non-small cell lung cancer

This study demonstrates that integrating patient-reported insomnia changes with a mathematical tumor growth model can predict treatment progression in non-small cell lung cancer patients with approximately 72% accuracy, enabling clinicians to intervene 6–8 weeks earlier than usual.

Original authors: Upadhyaya, D. J., Schabath, M. B., Hoogland, A. I., Brady-Nicholls, R.

Published 2026-01-29

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Upadhyaya, D. J., Schabath, M. B., Hoogland, A. I., Brady-Nicholls, R.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to predict the weather. You have a very sophisticated computer model that looks at wind speed, temperature, and pressure (this is like the mathematical model the doctors use to track tumor size). However, this model sometimes misses the sudden, unexpected storms that ruin a picnic.

Now, imagine you also ask the people living in the area, "How is your sleep?" (This is the Patient-Reported Outcome, or PRO). The researchers in this paper discovered that if a patient starts sleeping poorly, it's like the sky turning a dark, stormy gray before the rain actually starts.

Here is the story of what this paper found, broken down simply:

The Problem: Waiting Too Long to See the Storm

Lung cancer is a tough opponent. For many patients, doctors wait for a CT scan (a detailed photo of the inside of the body) to see if the cancer is growing. But by the time the scan shows the tumor getting bigger, it's often too late to change the treatment plan effectively. The doctors were looking at the "storm" only after it had already started raining.

The Old Tool: The Mathematical Map

The researchers used a special mathematical map called a Tumor Growth Inhibition (TGI) model. Think of this as a GPS that predicts how a car (the tumor) will move based on how fast it was going before.

What it did well: It was pretty good at predicting when the car was driving smoothly (stable disease).
Where it failed: It struggled to predict when the car was about to speed out of control (progression). It was like a GPS that says "you're fine" right before you hit a cliff.

The New Secret Ingredient: The Sleep Tracker

The researchers decided to add a new layer to their GPS: Insomnia. They asked patients, "How well are you sleeping?" every two weeks.

The Discovery: They found a strong link between bad sleep and the cancer getting worse. It's as if the patient's body is sending an early warning flare: "Something is wrong, even if the CT scan hasn't seen it yet."
The Result: By combining the GPS (the math model) with the Sleep Tracker (the patient's report), they could predict the "storm" much earlier.

How They Tested It

They looked at 80 patients with a specific type of lung cancer. They used a clever trick called "Leave-One-Out." Imagine you have 80 students in a class. To test your prediction skills, you take one student out, study the other 79, and try to guess what the one student will do. Then you put them back and pick a different one. They did this for every patient to make sure their method wasn't just a lucky guess.

The Results: Seeing the Future Earlier

When they combined the math model with the sleep data:

Accuracy: They got the prediction right about 71% of the time.
Speed: This was the biggest win. They could tell a patient was going to get worse 6 to 8 weeks earlier than they could by just waiting for the CT scan.
The Trade-off: They were slightly less perfect at predicting who would stay stable, but they became much better at catching who was about to get worse. In medicine, catching the bad news early is often considered more important so doctors can switch treatments before it's too late.

The Bottom Line

This paper doesn't claim to have a cure, but it offers a new way to listen to the patient. By treating the patient's own report of their sleep as a vital sign—just as important as a blood test or an X-ray—doctors can get a "heads up" that the cancer is changing its behavior.

Think of it like this: The CT scan is a rear-view mirror showing you where the car has been. The patient's report on their sleep is the windshield, showing you the road ahead before you even hit the bump. By using both, the doctors hope to steer the car away from danger sooner.

Technical Summary: Forecasting Patient-Specific Tumor Response Using Patient-Reported Outcomes in Non-Small Cell Lung Cancer

Problem Statement
Non-small cell lung cancer (NSCLC) remains the leading cause of cancer death, with a five-year survival rate below 20%. While biomarkers are essential for assessing treatment response, molecular prognostic biomarkers are often excluded from routine use due to tumor heterogeneity and irreproducible results. Consequently, there is a need for non-molecular prognostic biomarkers. Patient-Reported Outcomes (PROs), which quantify quality of life directly from the patient without external interpretation, have shown correlation with tumor volume dynamics. However, prior research has been limited; specifically, previous work demonstrated that changes in insomnia could predict whether a patient would progress with 77% accuracy, but only 45 days in advance. Crucially, this approach failed to predict when progression would occur and showed poor accuracy (28%) in predicting resistance. The core problem addressed is the lack of a framework that integrates PRO dynamics with mathematical tumor modeling to precisely forecast the timing of volumetric treatment progression, thereby allowing for earlier clinical intervention.

Methodology
The study analyzed data from 80 NSCLC patients undergoing immune checkpoint inhibitor therapy at the H. Lee Moffitt Cancer Center.

Data Collection: Patients completed biweekly PRO surveys assessing 27 symptoms on a 5-point Likert scale. Tumor volume was measured approximately monthly via CT scans, evaluated using Immunotherapy Response Evaluation Criteria in Solid Tumors (iRECIST). Progressive disease was defined as a 20% increase from the nadir of the sum of longest diameters (translating to a 72.8% volume increase).
Mathematical Modeling: A Tumor Growth Inhibition (TGI) model was utilized to describe therapy effects, incorporating exponential growth, treatment-induced decay, and the evolution of resistance over time. The model parameters included pre-treatment growth rate ( $\lambda$ ), initial treatment sensitivity ( $\gamma_0$ ), and the rate of resistance evolution ( $\epsilon$ ).
Model Calibration: Using a leave-one-out (LOO) cross-validation approach, the model was calibrated for all 80 patients. A nested optimization function determined that $\lambda$ should be patient-specific, while $\gamma_0$ and $\epsilon$ were uniform across the cohort. The Akaike Information Criteria (AIC) confirmed this parameterization.
Prediction Framework:
- Without PROs: Predictions were generated by sampling $\lambda$ values from a Cumulative Distribution Function (CDF) derived from the training set.
- With PROs: The framework was enhanced by stratifying the CDF based on changes in patient-reported insomnia. Three distinct curves were established: one for patients with increased insomnia (sampled for potential progression), one for decreased/stable insomnia (sampled for response), and a combined curve for patients with no change.
Analysis: A cutoff analysis using the Youden Index was performed to determine optimal thresholds for distinguishing responders from non-responders. Kaplan-Meier plots were employed to estimate the cumulative probability of progression over time.

Key Contributions

Integration of PROs and Mathematical Modeling: The study successfully integrated dynamic changes in a specific PRO (insomnia) with a TGI mathematical model to forecast tumor volume dynamics.
Temporal Precision: Unlike previous binary classification methods, this framework aims to predict when progression occurs, offering a timeline for clinical intervention.
Stratified Parameter Sampling: The introduction of a third CDF curve for patients with no insomnia change addressed the issue of disproportionately weighted resistance rates, refining the prediction algorithm.
Validation of Non-Molecular Biomarkers: The work provides evidence that PROs can serve as actionable, non-invasive prognostic biomarkers when combined with mechanistic modeling.

Results

Model Calibration: The calibrated TGI model effectively captured patient-specific dynamics. Figure 1B demonstrated that pre-treatment growth rates ( $\lambda$ ) were significantly higher in resistant patients compared to responsive patients ( $p = 4.77 \times 10^{-5}$ ).
TGI Model Alone: When using the TGI model without PROs, the overall accuracy for predicting response from volume points one to two was 66.25% (Sensitivity: 75%, Specificity: 65.28%).
TGI Model with PROs: Incorporating insomnia changes improved performance. For predictions from volume points one to two, the model achieved:
- Sensitivity: 80% (8 true positives, 2 false negatives).
- Specificity: 68.2% (58 true negatives, 27 false positives).
- Mean Area Under the Curve (mAUC): 0.78.
- Overall Accuracy: 71.3%.
- True Positive Rate: 72.2%.
Temporal Prediction: Kaplan-Meier analysis allowed for the identification of patients with lower survival probabilities (higher risk of progression) on a volume-to-volume basis, providing a timeline for potential intervention.

Significance and Claims
The paper claims that this innovative framework enables clinicians to predict precisely when disease progression will occur, offering an opportunity to intervene before the progression is confirmed by imaging. The authors emphasize that while the TGI model alone is effective at predicting responsive disease, it struggles with resistant disease; the incorporation of insomnia data remedies this limitation, significantly enhancing the ability to predict resistance. The study concludes that PROs, when coupled with mathematical modeling, serve as a powerful, minimally invasive tool for forecasting treatment response. The authors maintain a modest stance, acknowledging limitations such as the small, unbalanced dataset (10 progressive vs. 85 responsive events in certain intervals) and the need for future work to address tumor heterogeneity and model convergence issues.