Patient Matching Algorithm

The foundational dichotomy in matching philosophy. Deterministic matching relies on exact, rule-based comparisons of specific identifiers (e.g., SSN, MRN, exact name match). It is fast and transparent but brittle in the face of typos or missing data. Probabilistic matching uses statistical models, such as the Fellegi-Sunter framework, to assign a likelihood score to a match based on the agreement and disagreement of multiple weighted attributes. This approach tolerates real-world data entropy, such as nicknames ('Bill' vs. 'William') and transposed digits, but requires careful threshold tuning to balance false positives against false negatives.

A computational optimization technique that prevents the algorithm from comparing every record against every other record—an O(n²) impossibility at scale. Blocking partitions the dataset into mutually exclusive buckets based on a high-selectivity key, such as phonetic surname encoding or ZIP code. Only records within the same block are compared. Advanced implementations use locality-sensitive hashing (LSH) or sorted neighborhood methods with a sliding window to recover matches that span block boundaries, dramatically reducing computational cost without sacrificing sensitivity.

Algorithms designed to overcome orthographic variation in names. Phonetic algorithms like Soundex, Metaphone, and Double Metaphone encode names by their pronunciation in English, collapsing 'Smith' and 'Smyth' to the same code. Fuzzy string metrics such as Levenshtein edit distance, Jaro-Winkler, and cosine similarity on character n-grams quantify the textual difference between two strings. These are critical for linking records across systems with inconsistent data entry standards, handling common misspellings, and bridging maiden-to-married name transitions.

A core probabilistic mechanism where not all demographic fields are treated equally. The algorithm assigns an agreement weight (log-likelihood ratio) to each field based on its discriminatory power and reliability. A match on Social Security Number carries a massive positive weight, while a match on gender carries a very low weight. Conversely, a mismatch on a highly reliable field like date of birth generates a large disagreement penalty. This weighting is often derived from the observed frequency of values in the population (e.g., a rare surname match is more significant than a common one like 'Smith'), allowing the composite score to reflect true match probability.

The process of tuning the match/no-match decision boundary to optimize for a specific business objective. A threshold curve is generated by running the algorithm on a labeled ground-truth dataset of known matches and non-matches. The final threshold is set by analyzing the trade-off between precision (avoiding false links that merge two distinct patients) and recall (avoiding false non-links that create duplicate records for the same patient). In healthcare, the cost of a false positive (a merged chart) is clinically catastrophic, so thresholds are typically calibrated to heavily favor high precision, often routing uncertain scores to a manual clerical review queue.

Cryptographic methods that allow patient matching across organizations without sharing identifiable plaintext data. Hashing transforms identifiers into a one-way digest, but is vulnerable to dictionary attacks on common names. Advanced protocols use Bloom filters to encode attributes into a privacy-preserving bit array, allowing similarity comparison via set-based metrics like Dice coefficient without revealing the original tokens. Secure multi-party computation (SMPC) takes this further, allowing two parties to compute a match score on their combined data without either party revealing its private inputs to the other, a critical requirement for cross-entity research networks.