TL;DR
This systematic review dives into prognostic models designed to predict the recurrence of hepatocellular carcinoma (HCC) after thermal ablation. The key takeaway is that combining clinical and imaging data leads to better predictions than using either alone, but current models are still preliminary due to inconsistent definitions and a high risk of bias.
What Problem Does This Paper Address?
Hepatocellular carcinoma (HCC) is a major health concern, and thermal ablation (TA) is a common treatment. However, a significant challenge is predicting whether the cancer will return after TA, either locally or as a new tumor elsewhere in the liver. Early detection of patients at high risk for recurrence is crucial for tailoring treatment strategies and follow-up plans. While various models have been developed to predict this recurrence, there's a lack of clarity on which ones are most effective and reliable for clinical use. This paper addresses this by systematically reviewing and comparing existing prognostic models, highlighting their strengths, weaknesses, and the variability in their development and performance.
Methods
The researchers conducted a systematic review, searching major databases (PubMed, Web of Science, Cochrane, Embase) for studies that developed prognostic models for HCC recurrence after TA. They focused on models using both clinical and imaging data, with reported performance on a test set. Key aspects evaluated included:
- Model Performance: Metrics like AUC-ROC (Area Under the Receiver Operating Characteristic curve) and C-index were used to assess how well models discriminated between patients who would and wouldn't experience recurrence.
- Feature Extraction and Modeling Techniques: They analyzed how predictive factors were identified (e.g., radiomics, deep learning) and how the models were built (e.g., logistic regression, random survival forest).
- Risk of Bias and Applicability: The Prediction model Risk of Bias Assessment Tool (PROBAST) was used to critically appraise the quality of the included studies and the reliability of their findings.
Datasets
The review included 16 studies that developed a total of 39 prognostic models. All these models were developed using retrospective data, primarily from China and Korea. The paper does not mention the creation of any new benchmark datasets.
Key Insights and Ideas
- The Power of Combination: Models that integrated multiple types of predictors (clinical + imaging) consistently outperformed those relying on a single type. This suggests that a holistic view is essential for accurate prognostication.
- Inconsistency is King (Unfortunately): A major hurdle is the lack of standardized definitions for outcomes (like "early recurrence" or "local tumor progression") and inconsistent reporting of predictive parameters. This makes direct comparison between models very difficult.
- High Risk of Bias: The majority of the reviewed models suffered from a high risk of bias, often due to retrospective data collection, single-center studies, and issues with how data was analyzed or reported. This significantly impacts the reliability and generalizability of their findings.
- External Validation is Crucial (and Lacking): While some models showed promising internal validation, only a few were externally validated. This is a critical step for ensuring a model's performance holds up in different patient populations and clinical settings.
Key Results
- A total of 16 studies and 39 prognostic models were included in the review.
- Models integrating both clinical and imaging predictors generally performed better.
- For predicting early recurrence (ER), the nomogram developed by Huang et al. [25] showed good performance (AUC-ROC 0.83) with external validation, though on a small cohort.
- For late recurrence (LR), Ma et al. [23] developed a model with a C-index of 0.77, showing good calibration.
- For local tumor progression (LTP), a DL-radiomics model by Chen et al. [17] demonstrated excellent performance (C-index 0.87) with external validation.
- However, most studies had a high risk of bias, particularly in the analysis domain, and many lacked independent internal or external test sets.
Limitations and Future Work
The authors acknowledge several limitations:
- Heterogeneity: The included studies were highly diverse in terms of predictor types, modeling approaches, outcome definitions, and validation methods, which limited direct comparability and precluded a meta-analysis.
- PROBAST Tool Limitations: The PROBAST tool, while useful, may not fully capture biases unique to AI-based prediction models.
- Focus on Recurrence: The review exclusively focused on recurrence outcomes and did not consider other important clinical outcomes like post-procedural liver decompensation.
Future research should prioritize:
- Standardized Outcome Definitions: Clear and consistent definitions for endpoints are essential.
- External Testing: Rigorous external validation on independent cohorts is crucial for assessing generalizability.
- Transparent Development: Adherence to reporting guidelines like TRIPOD and clear documentation of methods are needed.
- Addressing Bias: Future studies must actively work to mitigate risks of bias in study design and reporting.
Until these challenges are addressed, the current models should be viewed as promising but preliminary tools.
Summary updated