Data transformation system for clinical budgets

A computer-implemented method executed by one or more processors for automated reconciliation of clinical trial budgets, the method comprising:

• a. receiving, from each of a sponsor budget source, a Medicare Coverage Analysis source, and a site charge master source, a plurality of procedural items, each item including a procedure name, a standardized procedure code, a cost value, and a billing category;
• b. normalizing the procedure names by tokenizing text, lowercasing, removing punctuation, expanding abbreviations using a stored synonym dictionary, and mapping procedure codes to canonical code families;
• c. generating, for each procedural item, a fixed-length vector embedding that encodes the normalized procedure name and the numerical attributes of the item by applying a trained text-embedding function and then normalizing each vector to unit length;
• d. constructing in memory an approximate nearest neighbor index over vectors of items from the site charge master source, the index being graph-based and parameterized to bound search fan-out and to provide sublinear retrieval of candidate neighbors;
• e. for each item from the sponsor budget source, querying the index to retrieve a ranked set of candidate site items by cosine similarity of the corresponding vectors;
• f. computing, for each candidate pair, a composite match score that includes a code-match indicator, a cosine similarity term, a numeric similarity term based on the difference between cost values, and a billing-consistency term derived from the billing categories, each term weighted according to a stored configuration, and determining that a candidate pair satisfies a similarity threshold when the composite match score meets or exceeds a calibrated threshold;
• g. forming a weighted bipartite graph whose left nodes correspond to items from the sponsor budget source and whose right nodes correspond to items from the site charge master source, whose edges connect pairs that satisfy the similarity threshold, and whose edge weights equal the composite match scores;
• h. selecting a one-to-one assignment between sponsor items and site items by computing a maximum-weight matching over the weighted bipartite graph;
• i. for each matched pair, identifying a corresponding item in the Medicare Coverage Analysis source by either exact standardized procedure code equality or by applying the steps of elements (c) through (f) with the Medicare Coverage Analysis source serving as a comparison dataset, and adjusting the composite match score with a predetermined code-match weight when the Medicare Coverage Analysis code equals the sponsor code or the site code;
• j. updating a reconciled budget data structure in memory that stores, for each matched set of items across sources, a selected cost value equal to a maximum permitted value determined by a policy rule set that enforces Medicare Coverage Analysis allowances and contract caps, and that stores a resolved billing category determined by a rule engine, and that stores provenance fields identifying the source of the selected cost value;
• k. flagging as unresolved any sponsor budget item that remains unmatched after the maximum-weight matching and a permutation-based re-tokenization pass that reorders tokens and expands abbreviations, and storing for each flagged item the top candidate explanations and scores;
• l. Writing an audit record for each processed item that includes a vector fingerprint, the candidate list, the composite match score components and weights, a match or flag outcome, and the rule rationale applied during cost and billing resolution;
• and
• m. transmitting the reconciled budget data structure and the audit records as a machine-readable payload through an application programming interface to a budget management system.