Inference: Adaptive Questionnaire
Context
The Adaptive Questionnaire ranks possible observations for a selected target, then recomputes that ranking after each answer. It uses the current evidence, mutual information, and observation costs to guide a case-specific sequence.
We continue with the SC=0.5 model trained on 456 records. The demonstration starts with no evidence and a Malignant prior probability of 37.72%.
Usage
Select Menus > Inference > Adaptive Questionnaire.
Select diagnosis as the target, leave Based on a Target State off, enable Reset Evidence, and allow 20 alternative questions. This demonstration considers the full diagnosis distribution rather than a binary indicator for one selected target state.
Costs
Stop the questionnaire before changing costs. Open Menus > Edit > Edit Costs. Initially, every displayed node has cost 1. Changing concave points_mean to 0.01 moves it from the second question to the first when the questionnaire restarts. Costs can represent relative effort as well as money. Restore the cost to 1 before following the sequence below.
Adaptive Evidence-Seeking
The FNA image attributes in this dataset are obtained together. The following sequence illustrates the software’s evidence-ranking behavior; it is not a clinical test protocol.
With equal costs and no evidence, the target monitor appears first, followed by perimeter_worst, concave points_mean, concave points_worst, concavity_worst, concavity_mean, texture_worst, and fractal_dimension_mean. The native monitor exports below preserve the displayed order, read down the first column and then down the second.
Observation #1
Double-click the middle state of perimeter_worst: 101.65 < perimeter_worst ≤ 114.4. The monitor’s abbreviated label is ≤114.4; it denotes the interval above the preceding threshold.
The Malignant probability becomes 42.62%. The questionnaire moves the observed variable to the end and now ranks concave points_worst first among the remaining questions.
Observation #2
Observe the middle state of concave points_worst: 0.144 < concave points_worst ≤ 0.176. The Malignant probability increases to 83.19%. The remaining questions are reordered using both observations.
Observation #3
Observe the first state of texture_worst: texture_worst ≤ 23.3. The Malignant probability decreases to 45.75%.
Observation #4
Observe the middle state of concavity_worst: 0.208 < concavity_worst ≤ 0.366. The Malignant probability rises to 82.58%.
Observation #5
Observe the first state of fractal_dimension_mean: fractal_dimension_mean ≤ 0.056. The Malignant probability becomes 98.43%. Two connected measurements remain unobserved: concave points_mean and concavity_mean. We end this illustrative sequence here; it is not an exhaustive observation of every predictor.
Stop the questionnaire and remove all observations to return to the baseline prior of 37.72% Malignant and Joint Probability 100%. Observation costs remain at 1.
Workflow Animation
The following legacy animation uses an earlier model. Its variables and numerical sequence differ from the updated example above.
Further Use
The questionnaire recomputes the next question for the evidence currently available. A WebSimulator can make a configured questionnaire available to users through a browser. A Target Interpretation Tree provides another way to inspect an evidence sequence.