We use a supervised learning technique to carefully prepare a single memristor model to predict whether the patient has the sepsis. A memristor behaves as a resistor, with a (mem)resistance R that changes in the interval Rmin<R
R*. The procedure is repeated for every row in the patient data table. To optimize the weights, we use a genetic algorithm in combination with a gradient descent technique. Preliminary results with only a one-memristor system and a limited number of free parameters (to speed up the training) give the normalized utility score equal to 0.31 on the training data. From our experience with neuromorphic applications of memristor networks we are confident that the score can be significantly increased by considering more complex memristor networks and weight structures. Our model can be used to identify which clinical variables significantly contribute to sepsis by analyzing the size of the optimized weight coefficients.