Underdetection

Underdetection is the phenomenon in which a system or process fails to identify all relevant instances of a target event, resulting in fewer detections than the true number. It can occur in any domain that relies on sensors, measurements, or judgments to signal occurrences, including medical imaging, environmental monitoring, surveillance, fraud detection, and cybersecurity. Underdetection is distinct from overdetection, in which too many non-target events are flagged as detections.

Causes include limited detector sensitivity, high decision thresholds, low sampling frequency, and data quality issues such as noise or missing data. Model bias, imbalanced datasets, nonstationary targets, and calibration errors can also increase underdetection. In practical systems, hardware limitations, privacy constraints, and deliberate avoidance by evaders may create blind spots that contribute to underdetection.

Evaluation typically uses ground truth comparisons and metrics such as recall (sensitivity). Underdetection manifests as low recall and may be difficult to detect if the true rate is low. Validation often requires controlled benchmarks, annotation, or simulated data to quantify missed detections and to distinguish underdetection from occasional misses due to randomness.

Implications include missed opportunities for intervention, underestimated risk, biased statistics, and reduced system trust. In safety-critical settings, undetected events can have serious consequences. In research and policy, underdetection can distort prevalence estimates and hinder evidence-based decisions.

Examples include radiology systems missing small tumors, wildlife cameras failing to record rare species, fraud detectors overlooking novel schemes, and intrusion detection systems missing stealthy intrusions.

Mitigation strategies emphasize improving sensing coverage and quality, lowering thresholds where appropriate, and using complementary methods such as data fusion, anomaly detection, ensemble models, and human oversight. Regular calibration, robust validation, and budgeted trade-offs between false positives and false negatives are also important. Transparent reporting of detection rates helps stakeholders assess risk and reliability.