A Neural Network Approach to Treatment Optimization

An approach for optimizing medical treatment as a function of measurable patient data is analyzed using a two-network system. The two-network approach is inspired by the field of control systems: one network, called a patient model (PM), is used to predict the outcome of the treatment, while the other, called a treatment network (TN), is used to optimize the predicted outcome. The system is tested with a variety of functions: one objective criterion (with and without interaction between treatments) and multi-objective criteria (with and without interaction between treatments). Data are generated using a simple Gaussian function for some studies and with functions derived from the medical literature for other studies. The experimental results can be summarized as follows: 1) the relative importance of symptoms can be adjusted by applying different coefficient weights in the PM objective functions. Different coefficients are employed to modulate the tradeoffs in symptoms. Higher coefficients in the cost function result in higher accuracy. 2) Different coefficients are applied to the objective functions of the TN when both objective functions are quadratic, the experimental results suggest that the higher the coefficient the better the symptom. 3) The simulation results of training the TN with a quadratic cost function and a quartic cost function indicate the threshold-like behavior in the quartic cost function when the dose is in the neighborhood of the threshold. 4) In general, the network illustrates a better performance than the quadratic model. However, the network encounters a local minima problem. Ultimately, the results indicate a proof of idea that this framework might be a useful tool for augmenting a clinician's decision in selecting dose strengths for an individual patient need