TL;DR
This paper introduces a novel deep-transfer-learning model that leverages natural language processing (NLP) on serial free-text rectal MRI reports to predict patient survival. This approach demonstrates the potential to extract valuable prognostic information from unstructured medical data, outperforming traditional tumor markers.
What Problem Does This Paper Address?
Electronic Medical Records (EMRs) are a treasure trove of patient information, but a significant portion of this data exists as unstructured free-text clinical notes, such as radiological reports. Analyzing this unstructured data for clinical decision-making is challenging due to the lack of standardized formats. While NLP has been applied to EMRs, previous methods often lacked the sophistication to extract nuanced prognostic information crucial for patient survival prediction. This paper aims to bridge this gap by developing a method to effectively utilize serial radiological reports to predict rectal cancer patient survival, thereby unlocking the potential of this underutilized data.
Methods
The core of this research is a deep-transfer-learning model designed to predict patient survival. The process involves several key steps:
- Word Embedding: Radiological reports are first converted into embedding vectors using a pre-trained language model, specifically ClinicalBERT. This model leverages knowledge transferred from public clinical notes, allowing it to understand medical terminology.
- Feature Aggregation: For patients with multiple serial MRI reports, a Recurrent Neural Network (RNN) with an attention mechanism is used to aggregate the word embeddings from each report into a single patient-specific feature vector. This captures the evolution of findings over time.
- Survival Prediction: A Cox-proportional hazards (Cox-PH) model then uses this patient feature vector to predict survival risk.
The researchers also explored feature extraction through N-gram clustering to identify high- and low-risk textual patterns within reports and used t-distributed Stochastic Neighbor Embedding (t-SNE) for visualizing patient embedding vectors to understand model behavior.
Datasets
The study utilized a retrospective cohort of 4,338 rectal cancer patients diagnosed between April 2012 and October 2019 at Yonsei Cancer Center, Seoul, Korea. The dataset included rectal magnetic resonance imaging (MRI) reports, along with basic demographic and clinical data such as sex, age, and carcinoembryonic antigen (CEA) levels. The data was split into a training set (n=3470) and a test set (n=868). Notably, the model was trained using only the free-text radiological reports and survival/death events, excluding other clinical data to specifically evaluate the NLP approach.
Key Insights and Ideas
- Unlocking Unstructured Data: The most significant insight is that valuable prognostic information for rectal cancer survival can be reliably extracted solely from free-text radiological reports, without needing structured clinical data.
- The Power of Serial Data: Incorporating serial MRI reports, rather than just a single report, significantly improves survival prediction accuracy. This highlights the importance of tracking disease progression and treatment response as documented in sequential reports.
- ClinicalBERT's Advantage: Using a pre-trained language model like ClinicalBERT, which is trained on clinical notes, is more effective for this task than general-purpose language models. This suggests that domain-specific pre-training is crucial for medical NLP applications.
- Outperforming Traditional Markers: The deep-learning model's predictions showed a stronger correlation with actual survival time than the carcinoembryonic antigen (CEA), a commonly used tumor marker, and achieved higher accuracy in predicting N-year survival.
- Interpretable Features: The model's visualization capabilities allowed researchers to identify specific textual patterns (N-grams) within reports that are strongly associated with worse or better prognoses, offering clinical interpretability.
Key Results
- The deep-transfer-learning model achieved a higher C-index (a measure of survival prediction accuracy) when using serial MRI reports compared to using only a single report (0.595 vs. 0.579).
- Using two serial reports further improved the C-index to 0.667, demonstrating a clear benefit from temporal data.
- The model successfully stratified patients into high-, intermediate-, and low-risk groups, with statistically significant differences in survival between these groups (log-rank p-value < 2.0×10⁻³).
- The model's predicted risk showed a stronger correlation with survival time (R = -0.590) than CEA (R = -0.195).
- Receiver operating characteristic (ROC) curves indicated that the model's predicted risk yielded higher Area Under the ROC (AUROC) values for N-year survival prediction compared to CEA. For example, the 1-year survival AUROC was 0.804 for the model versus 0.557 for CEA.
Limitations and Future Work
- Single Institution Study: The study was conducted at a single institution, which might limit the generalizability of the model to other healthcare settings. However, the authors note that the model's robustness was demonstrated across reports from 25 radiologists, suggesting potential for adaptation with retraining.
- Retrospective Nature: The study design was retrospective, which is a common limitation in medical research.
- Future Directions: The authors suggest extending the model to other types of unstructured EMR data (e.g., clinical progress notes, dialogue records) and applying it to different cancer types. They also propose improving the survival-graph generation algorithm by incorporating additional clinical factors. The model's architecture itself is not specific to rectal cancer or MRI, indicating broad applicability.
Summary updated