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

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.

研究目的：遗传算法（Genetic Algorithm，GA）是一类机器学习优化策略，其通过样本策略间的适应度竞争迭代演化得到最优解。本研究借助遗传算法对中老年男性症状量表（Aging Male Symptoms，AMS）进行演化优化，以期通过血清睾酮水平更精准地识别迟发性性腺功能减退症（Late Onset Hypogonadism，LOH）。

材料与方法：本研究基于一项全国性迟发性性腺功能减退症流行病学研究中的标准中老年男性症状量表训练集，对遗传算法开展训练。针对量表特定条目生成随机选择器矩阵，每一代策略均通过以敏感度为基准构建的适应度函数完成演化。用于判定迟发性性腺功能减退症阳性的血清睾酮阈值，在每个竞争策略中均设置为随机值。在完成2000次迭代运行后，每一次运行均从3000组经300代演化得到的随机生成集合中选出最优结果；最终将得到的最优中老年男性症状量表选择矩阵，应用于独立招募的验证集以对比评估模型效果。

结果：预训练阶段中，当血清睾酮浓度高于3.5 ng/mL时，模型对睾酮水平的预测能力显著下降。将训练阶段的睾酮阈值限定在2.5~3.5 ng/mL区间内时，遗传算法演化得到93种不同的筛选策略。仅包含5个条目的筛选策略（设定评分为20分的阈值，且对应血清睾酮浓度为3.16 ng/mL）在内部验证集中展现出优异的重现性。将该筛选条件应用于独立验证集后，模型敏感度从0.66提升至0.77，特异度则从0.07提升至0.19。

结论：借助遗传算法优化问卷条目筛选的方法，可显著提升中老年男性症状量表的应用效能。该方法可便捷推广至其他无法与生理标志物建立关联的问卷研究中。