---
{
"title": "Introducing Model builder in MDMbox 2610.0",
"description": "Train patient matching models from FHIR data without labeled pairs. Review decisions and use the same model for online and bulk matching.",
"date": "2026-11-06",
"author": "Health Samurai Team",
"reading-time": "4 min read",
"tags": [
"FHIR Tools",
"SQL on FHIR",
"MDMbox"
],
"utm-campaign": "feature",
"utm-content": "model-builder"
}
---
> For the complete documentation index, see [llms.txt](https://www.health-samurai.io/llms.txt).
> Use it to discover all available pages before guessing URLs.
---
A misspelled name or an old address can hide a duplicate Patient record. **Model builder in MDMbox 2610.0 learns matching parameters from your FHIR Patient data**, without labeled training pairs. You choose the fields, inspect the decisions and put the saved model to work.
In our tests, automatic models achieved **99.45-100% F1 across six synthetic and error-simulated linkage datasets without SSN**. Against a population of **10 million Patient records**, the median online `$match` response took **25 ms**.
## Build from your Patient data
Choose the Patient fields to compare, such as `name`, `birthDate` and `address`. Preview their values and check for missing data. A SQL on FHIR `ViewDefinition` captures your field selection so training and matching use the same information.
Select **Build model**. The builder chooses which records to compare and learns how to weigh their similarities and differences. It handles spelling variations and missing values and can account for relationships between fields, such as a postal code predicting a city.
For custom logic, you can [edit comparisons in JavaScript](/docs/mdmbox/matching-models#automatic-training-and-javascript-extensions), with an optional AI assistant to help write them.
## See why records match
In **Review & save**, check the score distribution and open individual pairs. Each explanation shows how the selected fields contributed to the score, helping you investigate both likely matches and uncertain cases.
Adjust the minimum score for accepting a match or sending a pair for review. Then save a versioned model and use it for **online `$match`, Bulk matching and Continuous matching**. The field selection, learned parameters and thresholds stay together in the same model.
*The screenshots show synthetic Patient data.*
## Matching quality without labeled training pairs
We converted records from [FEBRL](https://users.cecs.anu.edu.au/~Peter.Christen/Febrl/febrl-0.3/febrldoc-0.3/node70.html), [ePBRN](https://github.com/ePBRN/Medical-Record-Linkage-Ensemble) and [NC Voters](https://dbs.uni-leipzig.de/research/projects/benchmark-datasets-for-entity-resolution) into **FHIR Patient resources** before training. The datasets' known identities let us measure precision, recall and F1 afterward. Those labels were kept out of training, and matches missed during candidate selection count against recall.
On the six smaller synthetic and error-simulated datasets, automatic training produced these results **without social security identifiers or handwritten comparison rules**:
| Dataset | Precision | Recall | F1 |
| --- | ---: | ---: | ---: |
| FEBRL1 | 100% | 100% | 100% |
| FEBRL2 | 100% | 99.49% | 99.75% |
| FEBRL3 | 100% | 98.91% | 99.45% |
| FEBRL4 | 100% | 99.53% | 99.77% |
| ePBRN D | 100% | 99.61% | 99.81% |
| ePBRN F | 100% | 99.28% | 99.64% |
Every accepted pair was correct in these tests. The test sets contain 158-1,924 known same-person pairs each.
### Matching against 5M and 10M records
The larger NC benchmarks use voter-derived names and address information with generated duplicates. We tested matching for **5,000 records per population**. Adjusting the acceptance thresholds on separate labeled examples recovered more true matches and raised F1 above **91.6%**, without retraining:
| Population | Setting | Precision | Recall | F1 |
| --- | --- | ---: | ---: | ---: |
| NC Voters 5M | Automatic | 95.58% | 73.91% | 83.36% |
| NC Voters 5M | Threshold adjusted | 91.56% | 92.03% | 91.79% |
| NC Voters 10M | Automatic | 95.37% | 76.45% | 84.87% |
| NC Voters 10M | Threshold adjusted | 92.23% | 91.13% | 91.68% |
Adjusting the threshold traded some precision for substantially higher recall. The downloads include the full quality results, including results for records with changed values.
## Millisecond matching at 10 million records
On an **Apple M1 Pro with 16 GB RAM**, automatic training took about **22 minutes for 5M Patient records** and **30 minutes for 10M**, after data preparation.
In online tests, median `$match` responses took **28 ms at 5M and 25 ms at 10M**. **95% arrived within 56 ms** on both populations. Training figures exclude data import and first-time online setup.
## Try Model builder
Open **Models → Model builder** in MDMbox `2610.0` to train your first model. Follow the [documentation](/docs/mdmbox/matching-models#model-builder-preview), or [book a demo](/mdmbox#contact-form) to discuss your Patient matching workflow with our team.
Try the same inputs using the [eight downloadable datasets](/docs/mdmbox/matching-models#benchmark-datasets). Each includes FHIR Patient NDJSON, its ViewDefinition and a separate complete identity-label CSV, including singletons. Labels are for external evaluation; they are not training inputs or an MDMbox label-upload format. The download documentation includes source attribution, terms and checksums.
For a small first trial, download FEBRL1: [Patient data](https://storage.googleapis.com/aidbox-public/mdmbox/benchmarks/20261106/febrl1/patients.ndjson.gz), [field mapping](https://storage.googleapis.com/aidbox-public/mdmbox/benchmarks/20261106/febrl1/viewdefinition.json) and [evaluation labels](https://storage.googleapis.com/aidbox-public/mdmbox/benchmarks/20261106/febrl1/ground-truth.csv.gz).
Download the measured models and full quality results for [FEBRL](/docs/mdmbox/assets/benchmarks/febrl-20261105.zip?v=15beb3abbb392726), [ePBRN](/docs/mdmbox/assets/benchmarks/epbrn-20261105.zip?v=6deb8f5832c30b86) and [NC Voters](/docs/mdmbox/assets/benchmarks/nc-voters-20261105.zip?v=16337bdaeec52a43). The archives contain model definitions and results, without source Patient data.