Artificial Neural Network for Prediction of Clinical Outcomes


A machine learning model to predict the need for ventriculoperitoneal shunt after posterior fossa tumor resection.

Key Benefits

  • Highly accurate prediction of post-surgical complications.
  • Reducing healthcare costs by predicting prognoses and increasing efficiency of post-surgical treatment.

Market Summary

Neurosurgical procedures completed in the posterior fossa are usually well-tolerated. However, a well-known complication after resection of posterior fossa tumors (PFTs) is post-operative hydrocephalus (HCP). Post-operative HCP can necessitate long-term CSF diversion, which neurosurgeons usually treat with a ventriculoperitoneal shunt (VPS). Researchers at Emory University designed a multi-layer perceptron (or ANN) model, using the data mining tool, WEKA, that includes 17 input variables with 2 hidden layers to predict the two output classes: VPS requirement vs. no VPS requirement. The model predicts, with high accuracy, the subset of patients who will likely require VPS placement post-operatively. The model will contribute to the diagnosis segment of the global artificial intelligence in cancer market, which is expected to continue growing.

Technical Summary

The invention consists of a machine-learning model developed with multi-layer perceptron (ANN) using the data mining tool WEKA that predicts the subset of patients who are a risk of requiring ventriculoperitoneal shunt (VPS) treatment post-operation. The model consists of 17 input variables with 2 hidden layers, which predict two output classes: VPS requirement vs. no VPS requirement. To test its performance the researchers trained the model using N=518 observations with 10-fold cross validation at a learning rate of approximately 0.05 using a batch size of 100. Model performance was determined against the weighted area and accuracy in predicting VPS class using the receiver operating characteristic curve (AUC). The results found that the positive predictive value was 83% and the native predictive value was 98.8%

Patent Information

Tech ID: 21176
Published: 11/27/2023