Evolutionary Design of Reduced Precision Levodopa-Induced Dyskinesia Classifiers

Cartesian genetic programming, Coevolution, Adaptive size fitness predictors, Energy-efficient, Hardware-oriented, Fixed-point arithmetic, Levodopa-induced dyskinesia, Parkinsons disease

Parkinson's disease is one of the most common neurological conditions whose symptoms are usually treated with a drug containing levodopa. To minimise levodopa side effects, i.e. levodopa-induced dyskinesia (LID), it is necessary to correctly manage levodopa dosage. This article covers an application of cartesian genetic programming (CGP) to assess LID based on time series collected using accelerators attached to the patient's body. Evolutionary design of reduced precision classifiers of LID is investigated in order to find a hardware-efficient classifier together with classification accuracy as close as possible to a baseline software implementation. CGP equipped with the coevolution of adaptive size fitness predictors (coASFP) is used to design LID-classifiers working with fixed-point arithmetics with reduced precision, which is suitable for implementation in application-specific integrated circuits. In this particular task, we achieved a significant evolutionary design computational cost reduction in comparison with the original CGP. Moreover, coASFP effectively prevented overfitting in this task. Experiments with reduced precision LID-classifier design show that evolved classifiers working with 8-bit unsigned integer data representation, together with the input data scaling using the logical right shift, not only significantly outperformed hardware characteristics of all other investigated solutions but also achieved a better classifier accuracy in comparison with classifiers working with the floating-point numbers.