Intra-arterial therapies are the standard of care for patients with hepatocellular carcinoma who cannot undergo surgical resection. A method for predicting response to these therapies is proposed. The technique uses pre-procedural clinical, demographic, and imaging information to train machine learning models capable of predicting response prior to treatment.
This method can help predict response to intra arterial therapies. Machine learning applications in interventional oncology will change the way we treat liver cancer and other diseases with image-guided therapies. This method has the potential to improve the way two important decisions are made.
Generally, individuals new to this method will struggle because of the need to generate the features and to perform machine learning and code. Visual demonstration of this method is critical as the feature generation steps entail extracting those features using imaging masks. A well-annotated and clinically representative dataset is necessary to train the model.
More patient data and large cohorts will be necessary to further improve such a machine learning algorithm. After deciding which clinical features to include in the model, use an appropriate natural language tool kit to parse the plain-text clinic notes into sentences to allow searching for terms of interest. Then store each patient's features in a file with one feature per line.
For non-binary features, obtain the median value of each feature across all of the patients, and binarize each feature as a true or false value based on the median value. To calculate a mean liver enhancement feature from a medical image. After isolating the voxels containing the liver, enter the mean liver enhancement command.
To determine the liver volume feature, enter the commands as indicated. Then calculate median values for each imaging feature and binarize the features as demonstrated. For aggregation and reduction of the features, first remove the low-variance features from consideration and read the features in the binary matrix.
Next, use the variance threshold model to compute the features appearing in at least 20%of both responders and non-responders, similar to that illustrated in the table. Remove any features with a low univariate association with the outcome and filter only those features that pass the original low variant screening retaining n features where n is the number of patients. Then read in the binary matrix for each feature, similar to as illustrated in the table.
After the features have been binarized and filtered, it's time to train the model. Model training is a straightforward process that fits the features to the outcome under study and can be used to make predictions about new patients. In this representative analysis of 36 patients who underwent trans arterial therapies for hepatocellular carcinoma, 25 features were identified and binarized as demonstrated with 5 features satisfying both the variants and the univariate association filters.
Each patient was labeled as either a responder or non-responder under the quantitative European Association for the study of the liver response criteria. As illustrated, both the logistic regression and random forest models predicted a trans arterial chemoembolization treatment response with an overall accuracy of 78%Following this method, it is possible to make predictions on new patients by applying the training model. This machine learning technique can be universally applied to any therapy that involves pre-procedural, inter-procedural, and post-procedural imaging.
