Decision Tree

A decision tree is a graphical and mathematical model used in clinical decision analysis and health economics to represent a sequence of decisions and chance events, mapping out possible pathways and their associated probabilities, costs, and outcomes. It takes the form of a branching diagram that starts from a single decision node and extends through chance nodes (where outcomes are determined by probability) to terminal nodes (where final outcomes are recorded). Decision trees are one of the simplest and most widely used decision modeling structures.

Clinical and policy decisions involve choosing among options with uncertain outcomes, and decision trees provide a transparent, structured way to think through those choices. By laying out all possible pathways and attaching probabilities and values to each, a decision tree makes the logic of a decision explicit and quantifiable. Clinicians use decision trees to compare diagnostic and treatment strategies. Health economists use them to estimate the cost effectiveness of interventions. Guideline developers use them to structure the evidence underlying clinical recommendations. The visual clarity of the tree format makes decision trees accessible to stakeholders who may not be familiar with more complex modeling techniques.

A decision tree begins with a decision node, represented as a square, where the decision maker faces a choice between two or more strategies (for example, treat with drug A, treat with drug B, or watchful waiting). Each branch from the decision node leads to a chance node, represented as a circle, where nature determines what happens next according to specified probabilities. The branches emanating from a chance node represent the possible outcomes (for example, treatment success or treatment failure), each with a probability that must sum to one across all branches from that node.

The process continues through as many levels of decisions and chance events as necessary to fully represent the clinical scenario. Each terminal node (represented as a triangle) records the final outcome value, which may be measured in clinical terms (life years gained, events averted), quality adjusted terms (QALYs), monetary terms (costs), or both.

The analysis works backward from the terminal nodes using a process called "folding back" or "rolling back" the tree. At each chance node, the expected value is calculated by multiplying each outcome value by its probability and summing across all branches. At each decision node, the strategy with the highest expected value (or lowest expected cost, depending on the analysis) is selected. This process identifies the optimal strategy under the given assumptions.

Sensitivity analysis tests how the optimal decision changes when input values (probabilities, costs, utilities) are varied. Threshold analysis identifies the exact value of a parameter at which the preferred strategy switches from one option to another. For example, a threshold analysis might show that drug A is preferred as long as its cure rate exceeds 65%, and drug B is preferred below that threshold.

Decision trees are best suited for decisions with a short time horizon and a manageable number of branches. As the number of possible events grows, trees become unwieldy (a phenomenon called "bushy tree" complexity), and Markov models or microsimulation approaches may be more appropriate.

Decision trees are appropriate for clinical decisions that play out over a short time horizon (days to weeks), involve a limited number of decisions and chance events, and can be adequately represented without modeling time dependent transitions between health states. Common applications include comparing diagnostic test strategies, evaluating first line treatment choices, and analyzing screening decisions. They are not appropriate for chronic diseases requiring modeling of recurring events over years or decades, for conditions with many possible health states, or for decisions where the timing of events matters. In these cases, Markov models, microsimulations, or discrete event simulations are more suitable.

A common error is building a decision tree that is overly simplified, omitting important clinical pathways or outcomes that could change the preferred strategy. The opposite mistake, creating an excessively complex tree with dozens of branches, makes the model difficult to understand, validate, and communicate. Assigning probabilities without adequate evidence, or using estimates from studies that do not match the patient population in question, undermines the credibility of the analysis. Failing to conduct sensitivity analysis, and presenting the tree's conclusions as certain rather than conditional on the input values, overstates the evidence. Finally, some analysts confuse decision trees used in clinical decision analysis (which compare strategies) with classification trees used in machine learning (which predict categories), applying methods from one context inappropriately to the other.

In a hypothetical example, a diagnostic decision tree compares two strategies for evaluating patients presenting with suspected pulmonary embolism: D dimer testing followed by CT pulmonary angiography (CTPA) only for those with elevated D dimer versus CTPA for all patients. The tree branches at the first chance node by D dimer result (positive or negative), then by CTPA result (positive or negative among those tested), and finally by true disease status (PE present or absent). Terminal nodes capture costs (laboratory test, CTPA, treatment, missed PE management) and quality adjusted life days. Folding back the tree shows that the D dimer first strategy dominates in low risk patients (lower cost and higher expected QALDs) but is inferior in high risk patients, where the higher prevalence of PE means that a negative D dimer result misses more cases (its negative predictive value is lower).