Predicting Survival in Patients with Colorectal Liver Metastases Using CT Imaging and Clinical Parameters: Introducing a User-Friendly Online Calculator

Technical Summary: Predicting Survival in Patients with Colorectal Liver Metastases Using CT Imaging and Clinical Parameters

Problem Statement
Colorectal cancer (CRC) frequently metastasizes to the liver, with 15% to 25% of patients presenting with colorectal liver metastases (CRLM) at diagnosis. While hepatic resection offers improved long-term survival (47%–60%), recurrence rates remain high, and the optimal timing for resection is still under investigation. A significant clinical challenge lies in identifying suitable candidates for surgery and accurately assessing postoperative prognosis. Although Computed Tomography (CT) is the standard imaging modality for detection and staging, existing prediction models often lack the integration of readily available clinical and radiological data into a user-friendly tool for individualized risk stratification. Furthermore, while machine learning (ML) is gaining traction, it often requires large datasets and lacks the transparency and accessibility of traditional regression models, particularly in resource-limited settings.

Methodology
The study employed a retrospective cohort design using data from 174 patients who underwent liver resection for CRLM at Memorial Sloan Kettering Cancer Center (MSKCC). The dataset was sourced from The Cancer Imaging Archive (TCIA).

• Data Collection and Variables: The study integrated preoperative clinical parameters and CT imaging findings. Predictors included demographic factors (age, sex, BMI), clinical history (major comorbidities, preoperative chemotherapy, portal vein embolization), and specific CT features (tumor size, multiplicity, bilobar disease, synchronous metastasis, extrahepatic disease, lymph node involvement). Pathological features such as total treatment response rate, hepatic sinusoidal dilation, and NASH scores were also incorporated.
• Statistical Analysis: Cox proportional hazards regression was utilized to develop two distinct models: one for Overall Survival (OS) and one for Disease-Free Survival (DFS/Recurrence). The study adhered to the TRIPOD guidelines.
  • Model Validation: Internal validation was performed using bootstrap and Jack-Knife resampling methods.
  • Performance Metrics: Model discrimination was assessed via time-dependent Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC). Calibration was evaluated using calibration plots and p-values. Goodness-of-fit was measured using the Akaike Information Criterion (AIC) and Brier scores.
  • Variable Selection: Dominance analysis was conducted to rank predictors based on their contribution to outcome classification. Collinearity was inspected using Variance Inflation Factors (VIF), and no substantial multicollinearity was found.
• Tool Development: Based on the validated models, a free, open-access, web-based calculator was developed to estimate personalized probabilities of death and recurrence at 3, 5, 8, and 10-year intervals.

Key Results

• Cohort Characteristics: The study included 174 patients, with 87 deaths (50.0%) and 108 recurrences (62.1%) observed during follow-up.
• Predictive Performance:
  • Overall Survival (OS) Model: Achieved an AUC of 0.700 (95% CI: 0.618–0.783) and a Brier score of 0.219. The model demonstrated good calibration (p=0.92).
  • Recurrence (DFS) Model: Achieved an AUC of 0.724 (95% CI: 0.645–0.802) and a Brier score of 0.203, with good calibration (p=0.72).
• Risk Stratification: Using optimal cut-off points derived from the Youden index (59.6% for OS and 33.5% for recurrence), patients were stratified into low-risk and high-risk groups.
  • The high-risk group for OS had a Hazard Ratio (HR) of 2.83 compared to the low-risk group, with a mean OS of 73.1 months versus 105.1 months.
  • The high-risk group for recurrence had an HR of 2.90, with a mean DFS of 38.0 months versus 80.7 months for the low-risk group.
• Key Predictors: Dominance analysis identified the most informative variables for long-term OS as: major comorbidity, maximum tumor size, preoperative chemotherapy, total response rate, sex, multiple metastases, and synchronous liver metastasis.
• Decision Curve Analysis: Both models showed a positive net benefit, particularly for patients with intermediate or indeterminate prognoses (25–85% probability of death/recurrence).

Key Contributions

1. Integrated Prognostic Models: The study successfully developed and validated prognostic models that combine easily accessible clinical data with standard CT imaging features, avoiding the need for complex radiomics or molecular markers.
2. User-Friendly Clinical Tool: The authors introduced a free, open-access online calculator that translates complex statistical models into actionable, personalized risk probabilities for clinicians.
3. Comparative Performance: The models demonstrated superior predictive accuracy (AUC 0.70 for OS, 0.72 for recurrence) compared to previously published machine learning models that relied on radiomic features (which reported AUCs of 0.62 and 0.59, respectively).
4. Accessibility: By utilizing traditional regression methods and standard CT parameters, the tool is designed to be interpretable and applicable in centers with limited resources or those lacking advanced imaging processing systems.

Significance and Claims
The authors claim that integrating clinical parameters and CT imaging into an accessible model allows for precise prediction of long-term OS and DFS in CRLM patients undergoing hepatic resection. They posit that incorporating these models into conventional evaluation protocols can improve risk stratification and treatment decision-making. The study emphasizes that the developed calculator aids in standardizing healthcare practices, reducing physician error in complex calculations, and facilitating shared decision-making between providers and patients. The authors note that while the models perform well, they are based on retrospective data and that future studies should explore the combination of these clinical parameters with machine learning and radiomics to potentially enhance treatment planning further. They also acknowledge limitations regarding the detection of sub-centimeter metastases by CT compared to MRI.