Bring Patient Data
Bring patient data turns any set of CSV files, or one Excel workbook with several tabs, into a dataset agents can read. Columns can be anything: you tell Medbot what each table means, and which column names the patient, and it joins the tables into one record per patient.
Open it from Datasets → Bring patient data (medbot.me/datasets/import ). The import has three steps: Choose files, Map columns and Review and save. Nothing is uploaded until you click Save as dataset.
The page’s other tabs point elsewhere: FHIR connector and Notes connector explain connectors and link to them, and FHIR JSON and other opens the other imports.
Step 1: Choose files
Drop files onto the page or click Choose files. You can bring .csv, .tsv and .xlsx files, up to 200 MB in total.
- Each CSV or TSV file becomes one table, named after the file.
- Each tab of an Excel workbook that has rows becomes a table, named after the tab.
- Blank rows, and columns without a header, are dropped.
Each table shows its row and column counts, its headers, and the kind of data Medbot thinks it holds (or Unsure, check). The patient column can be named differently in each table: rows are joined by the value, not the column name.
Click Continue to mapping. Over 100,000 rows in total, Medbot warns that the upload will be slow and suggests importing in parts.
Step 2: Map columns
For each table in the Tables list, say what the table describes and which column names the patient. Everything else is optional. Medbot guesses all of this from the headers and sample values; check its guesses. A table marked Needs a role has something to fix.
This table describes
| Kind | Becomes |
|---|---|
| Patients | Patient |
| Lab results | Observation |
| Medications | MedicationStatement |
| Problems and diagnoses | Condition |
| Allergies | AllergyIntolerance |
| Clinical notes | DocumentReference |
| Encounters | Encounter |
| Other | Observation, or a note when it has free text |
Changing the kind guesses the column roles again.
Patient identifier and date
- Patient identifier: the column that names the patient, such as an MRN or study id. Medbot shows how many patients the table has or, when there is a patients table, how many of this table’s patients it matches, and how many rows have no identifier.
- Date: the column that dates each row, and whether dates are written D/M/Y or M/D/Y. Dates can be ISO (2024-01-12), day/month/year or month/day/year with / or -, or written with a month name (12 Jan 2024). Medbot tells you how many values it could not read.
Column roles
Give each column a role:
| Role | Use it for |
|---|---|
| Patient identifier | The column that names the patient. One per table. |
| Name, Date of birth, Sex | Patient demographics (patients tables). |
| Date | When the row happened. One per table. |
| Name or code | What the row is: the test, drug, diagnosis or allergy. |
| Value, Unit | A result and its unit. |
| Status or flag | A status or abnormal flag. |
| Free text, kept as a note | Text kept verbatim as a note, which can be extracted later. |
| Ignore | Leave the column out of the patient record. |
The roles offered depend on the table’s kind. Lab results, medications, problems and allergies tables need a Name or code column, and clinical notes tables need a Free text column. Extra columns on a patients table become per-patient observations named after the column.
To reuse a mapping, click Apply the same roles to and choose another table; columns with the same header get the same role.
One patient, as an agent will read it
The right-hand panel assembles one patient across every table, as a timeline, newest first, exactly as an agent will read it. Pick a patient to check the join. Values are shown exactly as recorded: Medbot never rounds, abbreviates or re-cases clinical data.
Step 3: Review and save
Name the dataset (Dataset name; it defaults to the first file’s name). The tiles summarise what will be saved:
| Tile | What it shows |
|---|---|
| Patients | How many patients can be analysed. |
| Records attached | Rows attached to a patient, by kind. |
| Date range | The years the records cover, and the median number of records per patient. |
| Rows not attached | Rows kept aside rather than dropped. |
The patient table lists each patient with the tables they appear in, their counts, their latest record and a status: Analysable, No notes (there are notes tables but none for this patient) or No medicines. Click a patient to preview their record below.
Rows Medbot could not attach
- Rows with no patient identifier.
- Rows whose patient appears in no patients table (only when you have a patients table).
- Rows whose date could not be read, with a link to change the table’s date order.
Download rows saves the unattached rows as one CSV per table. Tick Create patients from them to add rows whose patient is not in a patients table as new patients.
Saving
Click Save as dataset. Patients tables are saved first, then each table in chunks, with a progress bar. When it finishes you go to the new dataset. Its source is Tables from files.
- Importing the same rows again updates them rather than duplicating them.
- A patients table’s demographics are never overwritten by later tables.
- Rows with nothing to keep (for example a lab row with no test name and no value, or a note with no text) are skipped.
To pull records out of free-text notes as well, click Extract from notes with a connector; see Notes extraction.