Prediction of Lung Metastasis from Hepatocellular Carcinoma using the SEER Database
Jeff J.H. Kim, George R. Nahass, Yang Dai, Theja Tulabandhula · 2025
Read Prediction of Lung Metastasis from Hepatocellular Carcinoma using the SEER Database by Jeff J.H. Kim, George R. Nahass, Yang Dai, Theja Tulabandhula (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-11720.txt.
Introduction
Jeff Kim and co-authors build machine learning models that predict whether liver cancer will spread to the lungs, in a 2025 paper. Lung spread is the most common distant site for this cancer and darkens the outlook sharply.
The team trained four model types on a large United States cancer registry. Random forest and a neural network scored best overall, catching most true cases. Yet all models flagged many patients wrongly, so precision stayed low. A custom training goal pushed the network to miss fewer cases, and a combined model caught still more. Surgery status, tumor stage, and follow-up length mattered most among the inputs.
Catalog record
| CREATOR | Jeff J.H. Kim, George R. Nahass, Yang Dai, Theja Tulabandhula |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-11720
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Prediction of Lung Metastasis from Hepatocellular Carcinoma using the SEER Database Jeff J.H. Kim, George R. Nahass, Yang Dai, Theja Tulabandhula Abstract Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality, with lung metastases being the most common site of distant spread and significantly worsening prognosis. Despite the growing availability of clinical and demographic data, predictive models for lung metastasis in HCC remain limited in scope and clinical applicability. In this study, we develop and validate an end-to-end machine learning pipeline using data from the Surveillance, Epidemiology, and End Results (SEER) database. We evaluated three machine learning models—Random Forest, XGBoost, and Logistic Regression—alongside a multilayer perceptron (MLP) neural network. Our models achieved high AUROC values and recall, with the Random Forest and MLP models demonstrating the best overall performance (AUROC = 0.82). However, the low precision across models highlights the challenges of accurately predicting positive cases. To address these limitations, we developed a custom loss function incorporating recall optimization, enabling the MLP model to achieve the highest sensitivity. An ensemble approach further improved overall recall by leveraging the strengths of individual models. Feature importance analysis revealed key predictors such as surgery status, tumor staging, and follow-up duration, emphasizing the relevance of clinical interventions and disease progression in metastasis prediction. While this study demonstrates the potential of machine learning for identifying high-risk patients, limitations include reliance on imbalanced datasets, incomplete feature annotations, and the low precision of predictions. Future work should leverage the expanding SEER dataset, improve data imputation techniques, and explore advanced pre-trained models to enhance predictive accuracy and clinical utility. 1 Introduction Hepatocellular carcinoma (HCC) ranks as the sixth most common malignancy globally and is a leading cause of cancer-related mortality, with an alarmingly low five-year survival rate of approximately 16%. [ 1 , 2 ] Despite advancements in treatment, the prognosis for HCC patients remains grim, particularly when distant metastases develop. 5-year relative survival rates for liver cancer with distant metastasis can go low as 2.5%. [ 3 ] Among the common sites for metastasis, the lungs are most frequently affected, accounting for over 70% of metastatic cases. [ 4 ] These metastases significantly diminish survival outcomes, underscoring the need for effective predictive models for quicker treatment. The Surveillance, Epidemiology, and End Results (SEER) database, a comprehensive registry of cancer cases in the United States, has been extensively utilized to explore prognostic factors and develop predictive tools for cancer metastases. [ 5 ] Leveraging machine learning (ML) methodologies, recent studies have successfully created predictive models for HCC metastatic patterns, including lymph node metastases. [ 6 , 7 ] However, predictive models tailored specifically for lung metastases in HCC patients remain unexplored. Prior efforts often focus on isolated prognostic factors without integrating advanced computational methods to optimize predictive accuracy. [ 8 , 9 , 10 ] This gap presents an opportunity to enhance clinical decision-making by identifying high-risk patients earlier, allowing for more targeted surveillance and timely interventions. In this study, we aim to develop and validate a robust ML-based predictive model for lung metastasis in HCC patients using the SEER database. By integrating demographic, clinical, and tumor-specific variables, this model seeks to address the limitations of traditional approaches, offering improved precision and clinical utility. Additionally, the findings from this work could contribute to personalized treatment strategies and risk stratification, ultimately improving outcomes for this high-risk population. 2 Methods 2.1 Study Population Figure 1: Flowchart illustrating the inclusion and exclusion criteria for the patient population studied, as extracted from the Surveillance, Epidemiology, and End Results (SEER) database. The data is sourced from the Surveillance, Epidemiology, and End Results (SEER) database. [ 11 ] The inclusion criteria are as shown in Figure 1: 1) HCC patients diagnosed from 2010 - 2021. 2) age between 20 to 84. 3) Primary site is the liver. 4) One primary cancer only. 5) Known cause of death and followup survival months. 6) Summary and stage known. 7) Known metastasis status to the bone, brain, or lung. 8) Survival months greater than 0. 2.2 Data Preparation Redundant features such as the American Joint Committee on Cancer (AJCC) staging grade were consolidated. All feature names were lower cased and non alphanumeric characters were removed. Categorical features were one hot encoded when appropriate, and binary features were mapped to 0 and 1 (e.g. lung_met). Missing data points were imputed using KNN imputation for numeric features. The minority class was oversampled using the Synthetic Minority Oversampling Technique [ 12 ] . Figure 2: Correlation heatmap displaying the relationships between clinical and demographic variables in the SEER dataset. The color scale represents the strength and direction of the correlation, with red indicating positive correlations and blue indicating negative correlations. The intensity of the color corresponds to the magnitude of the correlation coefficient, with values shown within the cells. 2.3 Machine Learning Three individual machine learning models were trained for binary classification: XGBoost, logistic regression, and random forest. For all machine learning models, optimal hyperparameters were identified via grid search. A train and test split of 80/20 was used in all experiments. Feature importances were evaluated for all models. For logistic regression, the absolute values of model coefficients were calculated, as these represent the magnitude of each feature’s contribution to the prediction. For XGBoost and random forest models, feature importances were determined using the inherent importance scores provided by the model. These scores were sorted in descending order to identify the top contributors. 2.3.1 Multi Layer Perceptron A three layer multilayer perceptron (MLP) was trained for binary classification. After the first and second hidden a 30% dropout layer was used. Batch Norm and ReLU was used after every hidden layer. The final output layer consisted of two nodes, and the most probable class was selected following a softmax layer. Softmax was intentionally chosen in the event classification of more classes becomes pertinent in the future. The Adam optimizer with a learning rate of .001 was used and the model was trained for 100 epochs. The model with the highest F1-score was used for downstream tasks. A combination of the cross entropy loss ( 2 ) and recall loss ( 1 ) was used for the objective function. In our recall loss formulation, \hat{y}_{i}=\sigma(z_{i}) (where z_{i} are the output logits) is the predicted positive class, y_{i} is the true label, and N is the total number of samples. \mathcal{L}_{\text{recall}}=\frac{1}{N}\sum_{i=1}^{N}(1-\hat{y}_{i})y_{i} (1) \mathcal{L}_{\text{CE}}=-\frac{1}{N}\sum_{i=1}^{N}\left[y_{i}\log(\hat{y}_{i})+(1-y_{i})\log(1-\hat{y}_{i})\right], (2) The cross-entropy loss was weighted according to the distribution of class labels in the training dataset. The recall loss was weighted by a hyperparameter \gamma , so the final objective used is: \mathcal{L}_{\text{Total}}=\mathcal{L}_{\text{CE}}+\gamma\mathcal{L}_{\text{recall}} (3) 2.4 Statistical Analysis For all machine learning models, the following metrics were evaluated: \text{Accuracy}=\frac{\text{TP}+\text{TN}}{\text{TP}+\text{TN}+\text{FP}+\text{FN}} , \text{Precision}=\frac{\text{TP}}{\text{TP}+\text{FP}} , \text{Recall}=\frac{\text{TP}}{\text{TP}+\text{FN}} , \text{F1-Score}=2\cdot\frac{\text{Precision}\cdot\text{Recall}}{\text{Precision}+\text{Recall}} , and AUROC , which is the area under the receiver operating characteristic curve. 2.5 Hardware All experiments were run on an AMD Ryzen 7 7800X3D 8-Core Processor CPU. 3 Experiments Below we describe all experiments performed for lung metastasis prediction. 3.1 Data Exploration We initially probed the SEER dataset to identify any correlations between features ( 2 ). Lung metastasis had a positive correlation with tumor size ( r = 0.22), T staging ( r = 0.12), and N staging ( r =