Optimizing Aging Male Symptom Questionnaire Through Genetic Algorithms Based Machine Learning Techniques

Abstract

Purpose: Genetic algorithm (GA) is a machine learning optimization strategy where sample strategies compete for fitness to evolve an optimum solution. This study evolves the Aging Male Symptoms (AMS) with GA to better identify late onset hypogonadism (LOH) with serum testosterone.

Materials and methods: GA was trained on a training set of standard AMS questionnaire on a nationwide LOH epidemiology study. Random matrices of selectors for particular items were generated. Each generation of was evolved through a fitness function determined by sensitivity. Threshold to determine positive serum testosterone level for LOH was randomized for each competing strategy. After 2,000 runs, with each run producing the best result out of a set of 3,000 randomly generated sets evolved through 300 generations, the best AMS selection matrix was then applied to a separately enrolled validation set to compare outcomes.

Results: Predictability for serum testosterone levels dropped markedly above 3.5 ng/mL during pilot training. Limiting the training to testosterone thresholds between 2.5 and 3.5 ng/mL the GA 93 different strategies. Only a selection of 5 items, determining for a threshold of 20 points and determining for a serum testosterone level of 3.16 ng/mL, showed robust reproducibility within the internal validation set. Applying these conditions to the independent validation set showed sensitivity improved from 0.66 to 0.77, with a specificity of 0.07 to 0.19, respectively.

Conclusions: GA method of selecting questionnaires improved AMS questionnaire significantly. This method can be easily applied to other questionnaires that do not correlate with physiological markers.

Keywords: Hypogonadism; Machine learning; Questionnaire design; Testosterone.

Fig. 1. Methods of crossover mutation.

Fig. 2. Overall outline of analysis.

Fig. 3. Serum testosterone threshold vs. sensitivity displayed during 1,000 pilot training cycles; darker circles showing Aging Male Symptoms (AMS) weight at 3, lighter circles showing AMS weight at 4. There was no difference between applying AMS weight, however, sensitivity drop-off was noticeable above threshold of 3.5 ng/mL.

Fig. 4. Results of genetic algorithm show initial selections which were contrasted against internal validation to discriminate final selection matrix.