This paper describes the development of efficient, convenient, and non-repetitive short assessments (e.g., surveys, tests, exams) to replace a long assessment for conducting any kind of assessments of people's behaviors, opinions, attitudes, mental and physical states, traits (e.g., personality), and abilities (e.g., IQ), and skills (e.g., music playing proficiency). The process involves (1) obtain data for a particular long assessment (e.g., a scale for Generalized Anxiety Disorder, GAD) from a large sample of participants, (2) derive the ground truth scores from participants' responses to all items of the long assessment (e.g., scores of GAD or class of GAD: with vs. with GAD), (3) use machine learning to select participants' response to a small set of items of the long assessment to predict the ground truth accurately, (4) obtain computational models that use the minimal number of items needed to achieve the prediction accuracy as high as when the responses to all items of the whole long assessment are used, (5) generate all possible combinations of minimal number of items to create multiple short assessments of similar predictive accuracies for use if the short assessment is to be done repeatedly, (6) implement the short assessments in computer programs, mobile apps, websites, clouds, telemarketing, and any other forms for use with any future participant of the assessment, (7) the implementation will randomly or non-randomly select one subset of items to assess a particular participant using these platforms, (7) the implementation will use the computational model based on the subset of items to produce estimated scores as if the participant has done the long assessment, and the confidence of the model's estimation in the form of likelihood ratio or similar indexes.
A version of this paper has been published in the Behavior Research Methods journal.
This study applies a set of machine learning methods to develop a shortened, 100-item version of the MMPI-2 scale and evaluate its reliability and validity. We first selected the most important items from the original 567-item MMPI scale using feature importance measures such as MRMR and SHAP. Then we trained machine learning models on the selected items using a novel machine learning method called the stacked generalization ensemble of several machine learning techniques. We were able to reduce the length of the MMPI-2 by over 82% from 567 items to 100 items. The shortened MMPI scales, denoted the MMPI-100, can measure the 10 target conditions with at least 85% Area Under Curve of the Receiver Operating Characteristic (AUC ROC). It also has a Cronbach's Alpha of 0.96. Finally, we implemented this shortened MMPI-2 scale into a mobile-friendly web application. This shortened scale can not only help alleviate pressure on healthcare systems for dealing with the increased number of patients with mental health concerns but also for individuals to use it for regular mental health monitoring or screening.