Who it is about, and what was ruled out: four open models for Spanish clinical notes

Community Article
Published November 22, 2026

Se descarta tromboembolismo pulmonar.

Pulmonary embolism, ruled out. Search a hospital's Spanish discharge notes for "tromboembolismo pulmonar" and the note this sentence comes from is a hit twice: once in the CT result, once in the final diagnosis. The patient never had one. Both times, the note says so.

And before anyone runs that search, the note has to stop saying who she is. Her name, age, record and social security numbers, street address, phone number, the dates of her stay, the hospital, her son's name and the signing doctor's name, license number and e-mail all sit in the same twelve lines.

Those are the two jobs we trained four new open models for:

Job Model Parameters Runs with
Hide who the note is about OpenMed GLiNER2.5 PII Multilingual v1 287.4M OpenMed SDK
OpenMed mmBERT PII Multilingual Base v1 307.5M transformers
Read what was ruled out or only suspected OpenMed mmBERT ClinicalScope Spanish Base v1 307.5M transformers
OpenMed GLiNER2.5 ClinicalScope Spanish Decide 1B v1 1.19B OpenMed SDK

All four are Apache-2.0 and run on CPU.

To show what they do, we wrote one Spanish discharge note for this article. Every name, number, address, e-mail and hospital in it is invented. We ran it once, on CPU: OpenMed GLiNER2.5 PII masked it, then both scope models read the masked text one sentence at a time, every word marked in turn. Here is the whole run, misses included.

The synthetic discharge note after the real run, typed on a form. Identifiers under slate tape marked PII: hospital, patient name, age, record number, social security number, address in three pieces, phone, discharge date, the son's relationship word and name, the doctor's name, license number and e-mail, and the word Madre. Negation scopes from OpenMed mmBERT ClinicalScope underlined solid with a NO tag: diabetes mellitus, alergias medicamentosas conocidas, se palpan adenopatías cervicales, edemas en miembros inferiores, defectos de repleción, tromboembolismo pulmonar twice, and puede descartarse una neoplasia pulmonar subyacente. Uncertainty scopes underlined dashed with a question mark: derrame pleural derecho mínimo and Probable enfermedad pulmonar obstructiva crónica no diagnosticada. Pen notes: the admission date 03/02/2025 was left visible; Madre was masked, so the cancer no longer says whose it is; neoplasia was read as ruled out, while Decide 1B read it as suspected.

The note follows the order of the form it is written on, and so does this article: the patient's details first, then the clinical judgement, then what the run got wrong, the code, and how the four models were tested.

Datos del paciente: who the note is about

On our note. One analyze_text call with the settings from the model card returned 16 spans in 0.41 seconds once the model was loaded. It masked the hospital, the patient's name, "67 años", the record number (NHC), the social security number (NASS), the street address, the postcode and the city, the phone number, the discharge date, the son's name, the doctor's name, his license number and his e-mail.

Two things it did that a reviewer has to see:

  • It missed the admission date. "Fecha de ingreso: 03/02/2025" stayed visible, while the discharge date in the same line was masked (0.99).
  • It masked "Madre" and "hijo", the words for mother and son (1.00 and 0.69). For the son, that costs nothing: his name was masked too. For the mother it changes the note. "Madre fallecida por cáncer de mama" became "[PII] fallecida por cáncer de mama", and the sentence no longer says whose breast cancer it was.

The mmBERT PII model, run on the same note with its card's line-by-line code, returned 15 spans. It caught the same names and numbers and also masked "Madre" and "hijo". It missed both dates, masked "Hospital Universitario Monte" and left "Albar" visible, and masked the job title "adjunto".

On 91 held-out case reports. Both models were scored once on 91 Spanish clinical case reports from the MEDDOCAN corpus, with 2,065 annotated identifiers, from 90 article groups kept out of training and tuning. Scoring counts characters in the original text. F2 weighs a missed identifier twice as heavily as an extra mask.

Character F2 Exact-span F1 Critical identifiers fully masked Reports with any identifier character left
Before fine-tuning: GLiNER2.5 base model, same prompt, threshold 0.5 35.77 37.77 51.27 91 of 91
Earlier OpenMed PII model (frozen reference, not released) 84.64 43.12 81.71 91 of 91
OpenMed mmBERT PII Multilingual Base v1, threshold 0.35 98.12 95.97 97.89 28 of 91
OpenMed GLiNER2.5 PII Multilingual v1, threshold 0.15 98.57 97.60 98.94 19 of 91

Against the earlier OpenMed model, character F2 rose by +13.92 points for GLiNER2.5 (95% interval +12.80 to +15.10) and +13.48 for mmBERT (+12.52 to +14.51), from a paired bootstrap over the 90 article groups with 2,000 resamples.

Of the 946 critical identifiers (names, ID numbers, e-mails, phone and fax numbers, relatives and street addresses), GLiNER2.5 fully masked 936 and mmBERT 926:

Waffle chart, one square per critical identifier, 946 in all, in 91 held-out Spanish case reports. Filled squares were fully masked, open squares still show at least one character. Before fine-tuning: 485 masked, 461 showing, 51.27%. Earlier OpenMed PII model: 773 masked, 173 showing, 81.71%. OpenMed mmBERT PII: 926 masked, 20 showing, 97.89%. OpenMed GLiNER2.5 PII: 936 masked, 10 showing, 98.94%.

By type, the misses gather in three places for both models: professions (GLiNER2.5 fully masked 1 of 6, mmBERT 0 of 6), institutions (13 and 9 of 17) and relatives (40 and 35 of 44). The more common identifiers, names, dates, record numbers and license numbers, sit between 93 and 100 percent.

Heatmap of the share fully masked by identifier type, for 19 types with at least five annotations, across four models. Highlights: patient name 97.8 before fine-tuning, 100 for both OpenMed models; date 0.0 before fine-tuning, 100 for both; patient ID or record number 51.4, 90.5, 93.3, 96.2; relatives 11.4, 2.3, 79.5, 90.9; institution 0.0, 17.6, 52.9, 76.5; profession 0.0, 0.0, 0.0, 16.7.

Neither model masked anything in 24 short clinical notes written with no personal data in them (12 Spanish, 6 German, 6 French).

Beyond Spanish clinical notes. Both models also learned from synthetic finance documents in seven languages. On 588 held-out ones (84 per language, from 12 document types kept out of training), macro character F2 is 76.95 for GLiNER2.5 and 74.04 for mmBERT. These documents are synthetic, so they show transfer across languages, not performance on real records.

Character F2 Spanish English French German Italian Dutch Swedish
Before fine-tuning 42.78 42.39 34.80 40.77 43.40 29.34 36.20
Earlier OpenMed PII model 65.16 77.28 72.02 65.25 70.42 47.34 59.41
OpenMed mmBERT PII 68.72 89.27 75.96 78.81 76.78 59.56 69.20
OpenMed GLiNER2.5 PII 75.16 91.46 78.94 80.90 82.11 60.66 69.40

Dutch is the weakest language for both. German names are the one view where neither model beats the earlier OpenMed model. Of 107 person names in German court decisions, GLiNER2.5 fully masked 61.68% and mmBERT 36.45%, against 63.55% for the earlier model. For German text, use GLiNER2.5, and review the names.

Waffle chart, one square per person name in German court decisions, 107 in all. Filled squares were fully masked, open squares still show part of the name. Before fine-tuning: 57 masked, 50 showing, 53.27%. Earlier OpenMed PII model: 68 masked, 39 showing, 63.55%. OpenMed mmBERT PII: 39 masked, 68 showing, 36.45%. OpenMed GLiNER2.5 PII: 66 masked, 41 showing, 61.68%.

Juicio clínico: what was ruled out, what is only suspected

On our note. The two scope models read the masked note sentence by sentence, with each card's own code: every word is marked in turn and labeled as inside a negation scope, inside an uncertainty scope, or outside both. The mmBERT model labeled every word of the masked note in 1.8 seconds; Decide 1B took 30.8 seconds on the same machine.

Both put diabetes, the allergies, the cervical lymph nodes, the edema, the filling defects and both mentions of pulmonary embolism inside a negation scope. Both put the small pleural effusion ("Posible derrame pleural") and the undiagnosed COPD ("Probable enfermedad pulmonar obstructiva crónica") inside an uncertainty scope. Both left the pneumonia and the consolidation outside, which is right: those are present. We also ran the mmBERT model on the unmasked note, and masking changed none of its labels: the sentences that held identifiers had no scope in them, and the clinical sentences had nothing to mask.

They split on one finding, and it is the one that matters most in this note:

No puede descartarse una neoplasia pulmonar subyacente, por lo que se solicita TC de control en ocho semanas.

A lung cancer cannot be ruled out, so a control CT is booked. Decide 1B put "neoplasia" in an uncertainty scope (1.00). The mmBERT model, the one with the higher test score, put it in a negation scope, with a score of 0.91. A cohort query that drops negated findings would drop this patient from a search for suspected lung cancer. On its model card, mmBERT reads the same pattern, "No puede descartarse una rotura", as uncertainty. On this sentence it does not.

The scoped sentences of the masked note, word by word, with two strips under each word: OpenMed mmBERT ClinicalScope on top, Decide 1B below. Both agree on every finding in No se palpan adenopatías cervicales, sin defectos de repleción, que descarta tromboembolismo pulmonar, Posible derrame pleural derecho mínimo, Se descarta tromboembolismo pulmonar and Probable enfermedad pulmonar obstructiva crónica. They split in three places: Decide 1B reads cuidador principal as uncertain; they split on no diagnosticada; and on No puede descartarse una neoplasia pulmonar subyacente, mmBERT reads the neoplasia as ruled out and Decide 1B as suspected.

Two smaller differences show in the figure. Decide 1B put "cuidador principal" (main carer) in an uncertainty scope, which is not a finding at all. And the models often disagree on the cue words themselves (Niega, Sin, No, Posible): Decide 1B tends to include them in the scope, mmBERT tends to leave them outside. Cue words were never scored in testing, so read the findings.

On 2,028 held-out words. Both scope models were tested once on 2,028 marked words (867 inside a negation scope, 165 inside an uncertainty scope, 996 outside both) in 1,277 held-out sentences of annotated Spanish clinical text. The test sentences come from two clinical specialties, cardiology and traumatology, kept out of training and tuning.

Macro F1 Uncertainty F1 Uncertainty recall
Base GLiNER2.5 Decide 1B, zero-shot with the same prompts 33.26
Lexical control that sees only the marked word 49.52
Supervised lexical reference: TF-IDF and logistic regression with six words of context on each side, same training words 80.40 65.49
OpenMed GLiNER2.5 ClinicalScope Spanish Decide 1B v1 91.83 84.83 83.03
OpenMed mmBERT ClinicalScope Spanish Base v1 93.42 87.10 81.82

Against the lexical reference, mmBERT gains +13.02 macro F1 points (95% interval +10.51 to +15.49) and Decide 1B +11.43 (+9.06 to +13.72), from a paired bootstrap over 1,274 sentence groups with 2,000 resamples. The control that sees only the marked word reaches 49.52: scope lives in the context, not in the word.

Uncertainty is the weaker class for both. It is also the rarest in the data.

Two confusion matrices, rows for what the annotators marked and columns for what the model said. OpenMed mmBERT ClinicalScope, macro F1 93.42: negation 843 right, 2 read as uncertainty, 22 as outside; uncertainty 135 right, 12 read as negation, 18 as outside; outside 966 right, 22 read as negation, 8 as uncertainty. OpenMed GLiNER2.5 ClinicalScope Decide 1B, macro F1 91.83: negation 825, 6, 36; uncertainty 7 read as negation, 137 right, 21 as outside; outside 27, 15, 954 right. The uncertainty row is outlined.

In the 107 test words from sentences that hold both a negation and an uncertainty scope, accuracy is 91.59 for mmBERT and 88.79 for Decide 1B, against 63.55 for the lexical reference. Mixed sentences are harder for both models, and far harder for the lexical reference. The note above shows what can still slip: a doubt phrased with a negation ("No puede descartarse") read as a negation.

Observaciones: what this run got wrong

One note is not a test. It is a way to see the kind of mistake each model makes. From this run:

  1. GLiNER2.5 PII left the admission date visible, and masked the discharge date in the same line.
  2. GLiNER2.5 PII masked "Madre", and with it the information that the breast cancer was the mother's, not the patient's.
  3. mmBERT PII missed both dates, masked only part of the hospital name ("Albar" left visible) and masked the job title "adjunto".
  4. mmBERT ClinicalScope read "No puede descartarse una neoplasia" as ruled out, with a score of 0.91.
  5. Decide 1B put "cuidador principal" in an uncertainty scope.
  6. Fragments get fewer masks than whole notes. The date line on its own, through the same call, came back with neither date masked, and so did the patient and date lines together. Inside the whole note, the discharge date was masked. Pass whole notes, not snippets.
  7. Scores do not flag the errors. 145 of the 182 words mmBERT ClinicalScope labeled, and 178 of 182 for Decide 1B, scored 1.00 after rounding. The wrong neoplasia label scored 0.91. Use scores to rank, not as probabilities.

Every one of these is why the output of all four models goes to a person before a note leaves your control.

Run it yourself

The whole run above, from the note to the masked text, in one call:

pip install -U "openmed[gliner2]" "transformers>=5.17" torch
from openmed import OpenMedConfig, analyze_text

note = """INFORME DE ALTA. Servicio de Medicina Interna, Hospital Universitario Monte Albar.
Paciente: Rosario Gil Navarro, 67 años, NHC 6074215, NASS 37 20485519 03.
Domicilio: calle del Almendro 14, 3º izquierda, 37008 Salamanca. Teléfono: 923 418 552.
Fecha de ingreso: 03/02/2025. Fecha de alta: 09/02/2025.
Motivo de ingreso: disnea de dos días de evolución y dolor torácico de características pleuríticas.
Antecedentes: hipertensión arterial en tratamiento con amlodipino. Niega diabetes mellitus. Sin alergias medicamentosas conocidas. Madre fallecida por cáncer de mama.
Exploración: temperatura 38,4 °C, saturación de oxígeno 91 %. Crepitantes en base pulmonar derecha. No se palpan adenopatías cervicales. Sin edemas en miembros inferiores.
Pruebas complementarias: radiografía de tórax con condensación en lóbulo inferior derecho. AngioTC de arterias pulmonares sin defectos de repleción, que descarta tromboembolismo pulmonar. Posible derrame pleural derecho mínimo.
Evolución: buena respuesta a ceftriaxona y azitromicina, afebril desde el tercer día. Acude a las visitas acompañada de su hijo, Javier Gil Ramos, que será su cuidador principal.
Juicio clínico: neumonía adquirida en la comunidad en lóbulo inferior derecho. Se descarta tromboembolismo pulmonar. Probable enfermedad pulmonar obstructiva crónica no diagnosticada, pendiente de espirometría.
No puede descartarse una neoplasia pulmonar subyacente, por lo que se solicita TC de control en ocho semanas.
Firmado: Dr. Andrés Calvo Ortega, médico adjunto, nº de colegiado 373711904. Correo: acalvo.mi@hmontealbar.es"""

result = analyze_text(
    note,
    model_id="OpenMed/GLiNER2.5-PII-Multilingual-v1",
    labels=["personally identifying information"],
    threshold=0.1,
    confidence_threshold=0.15,
    max_length=120,
    config=OpenMedConfig(use_medical_tokenizer=False),
)
masked = note
for e in sorted(result.entities, key=lambda e: e.start, reverse=True):
    masked = masked[: e.start] + "[PII]" + masked[e.end :]
print(masked)
INFORME DE ALTA. Servicio de Medicina Interna, [PII].
Paciente: [PII], [PII], NHC [PII], NASS [PII].
Domicilio: [PII], [PII] [PII]. Teléfono: [PII].
Fecha de ingreso: 03/02/2025. Fecha de alta: [PII].
Motivo de ingreso: disnea de dos días de evolución y dolor torácico de características pleuríticas.
Antecedentes: hipertensión arterial en tratamiento con amlodipino. Niega diabetes mellitus. Sin alergias medicamentosas conocidas. [PII] fallecida por cáncer de mama.
Exploración: temperatura 38,4 °C, saturación de oxígeno 91 %. Crepitantes en base pulmonar derecha. No se palpan adenopatías cervicales. Sin edemas en miembros inferiores.
Pruebas complementarias: radiografía de tórax con condensación en lóbulo inferior derecho. AngioTC de arterias pulmonares sin defectos de repleción, que descarta tromboembolismo pulmonar. Posible derrame pleural derecho mínimo.
Evolución: buena respuesta a ceftriaxona y azitromicina, afebril desde el tercer día. Acude a las visitas acompañada de su [PII], [PII], que será su cuidador principal.
Juicio clínico: neumonía adquirida en la comunidad en lóbulo inferior derecho. Se descarta tromboembolismo pulmonar. Probable enfermedad pulmonar obstructiva crónica no diagnosticada, pendiente de espirometría.
No puede descartarse una neoplasia pulmonar subyacente, por lo que se solicita TC de control en ocho semanas.
Firmado: Dr. [PII], médico adjunto, nº de colegiado [PII]. Correo: [PII]
  • labels=["personally identifying information"] is the prompt the model was trained with. Every span comes back with this one label: the model marks what to mask, not which kind of identifier it is.
  • threshold=0.1 with confidence_threshold=0.15 is the decoding behind the evaluation. Raise confidence_threshold for fewer, surer spans, at the cost of more identifiers left visible.
  • max_length=120 groups sentences into pieces of at most 480 characters, close to the windows the model was trained and tested on. use_medical_tokenizer=False keeps the model's own span boundaries.

Then the scope of the findings, with the mmBERT ClinicalScope card's code. Pass one sentence at a time:

import re

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo = "OpenMed/mmBERT-ClinicalScope-Spanish-Base-v1"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo).eval()


def mark(sentence, start, end):
    """The evaluated input: "Texto: " and the sentence, with the word between ⟦ and ⟧."""
    return "Texto: " + sentence[:start] + "⟦" + sentence[start:end] + "⟧" + sentence[end:]


def scope(sentence, words=None):
    """Label every word of the sentence (or only `words`): negation, uncertainty or outside scope."""
    found = [m for m in re.finditer(r"\w+(?:-\w+)*", sentence) if words is None or m.group() in words]
    texts = [mark(sentence, m.start(), m.end()) for m in found]
    with torch.inference_mode():
        probs = model(**tokenizer(texts, padding=True, return_tensors="pt")).logits.softmax(-1)
    return [(m.group(), model.config.id2label[int(p.argmax())], float(p.max())) for m, p in zip(found, probs)]


for sentence, word in [
    ("Se descarta tromboembolismo pulmonar.", "tromboembolismo"),
    ("Posible derrame pleural derecho mínimo.", "derrame"),
    ("Probable enfermedad pulmonar obstructiva crónica no diagnosticada, pendiente de espirometría.", "enfermedad"),
    ("No puede descartarse una neoplasia pulmonar subyacente, por lo que se solicita TC de control en ocho semanas.", "neoplasia"),
    ("[PII] fallecida por cáncer de mama.", "cáncer"),
]:
    for w, label, score in scope(sentence, words={word}):
        print(f"{w:<16} {label:<17} {score:.2f}")
tromboembolismo  negation_scope    1.00
derrame          uncertainty_scope 1.00
enfermedad       uncertainty_scope 1.00
neoplasia        negation_scope    0.91
cáncer           outside_scope     1.00

The fourth line is the miss from the run. The last line is right, and still needs a reader: outside a scope does not mean the patient has it, and here the masked word was "Madre".

The same sentence through Decide 1B, with its card's code:

import re

from openmed.ner.families.gliner2 import load_gliner2_handle

handle = load_gliner2_handle("OpenMed/GLiNER2.5-ClinicalScope-Spanish-Decide-1B-v1")

# The task name and the three prompts the model was trained with
TASK = "Negation or uncertainty scope of the marked Spanish clinical term"
PROMPTS = {
    "negation scope: the marked term is under negation in the sentence": "negation_scope",
    "uncertainty scope: the marked term is under uncertainty in the sentence": "uncertainty_scope",
    "outside scope: the marked term is outside both scopes in the sentence": "outside_scope",
}


def mark(sentence, start, end):
    """The evaluated input: "Texto: " and the sentence, with the word between ⟦ and ⟧."""
    return "Texto: " + sentence[:start] + "⟦" + sentence[start:end] + "⟧" + sentence[end:]


def scope(sentence, words=None):
    """Label every word of the sentence (or only `words`): negation, uncertainty or outside scope."""
    out = []
    for m in re.finditer(r"\w+(?:-\w+)*", sentence):
        if words is None or m.group() in words:
            result = handle.classify_text(mark(sentence, m.start(), m.end()), {TASK: list(PROMPTS)})[TASK]
            best = max(result.scores, key=lambda s: s.score)
            out.append((m.group(), PROMPTS[best.label], best.score))
    return out


sentence = "No puede descartarse una neoplasia pulmonar subyacente, por lo que se solicita TC de control en ocho semanas."
for w, label, score in scope(sentence, words={"neoplasia"}):
    print(f"{w:<16} {label:<17} {score:.2f}")
neoplasia        uncertainty_scope 1.00

The mmBERT PII model runs with plain transformers; its card has the short find_pii function and the line-by-line version for long notes. Each model card has ten or more examples with real outputs: GLiNER2.5 PII, mmBERT PII, mmBERT ClinicalScope and Decide 1B ClinicalScope. The OpenMed SDK is Apache-2.0.

Pruebas: how the four models were tested

The PII pair. Both models learned from 303 Spanish clinical case reports of the MEDDOCAN corpus and 2,730 synthetic finance documents, with every identifier type mapped to one generic mask. Thresholds were chosen on a separate tuning set of 79 case reports and 441 finance documents, to favor recall. The earlier OpenMed PII model used as the reference was chosen on the tuning data too, from three earlier OpenMed PII models. Then the models, thresholds and scoring code were frozen, and each test view was scored once:

  • 91 Spanish case reports from 90 article groups held out of training and tuning, with 2,065 identifiers, 946 of them critical. The case text is authentic; the identifiers in it were substituted or added by the corpus authors. All of them come from the corpus authors' training file, not the official MEDDOCAN test set, so these numbers are not comparable to published MEDDOCAN results. 27 case reports linked to earlier OpenMed Spanish data were left out.
  • 588 synthetic finance documents in seven languages, from 12 document types held out of training. Dates, times and company names are outside the mask scope there.
  • 107 person names in German court decisions, counted as fully masked or not. The gold marks names only, so there is no precision on this view.
  • 24 short clinical notes with no personal data, written as controls, counted as masked if anything in them is masked.

The scope pair. Both models learned from 7,755 marked words in sentences from 14 clinical specialties of annotated Spanish clinical text, for two fixed epochs, and were checked on 2,258 marked words from two further specialties, internal medicine and urology. The 2,028 test words from cardiology and traumatology were scored once at the end, with the models, the references and the scoring code frozen first. The marked words are nouns and adjectives of three or more letters; the cue words themselves (no, sin, posible) are never marked. The source carries no patient identifiers, so the split is by specialty and sentence, not by patient. Macro F1 averages the F1 of the three classes.

The note in this article was written once, before any model ran, and was not edited after. Its run is saved with the hash of the note, the Hub revision of each model, the package versions and every span, label and score.

What they are not

  • Not a de-identification guarantee. Even on the in-distribution Spanish test, GLiNER2.5 left part of an identifier in 19 of 91 case reports, and the run above left a date visible. Have a person review the output before any text leaves your control, and follow the rules that apply to your data.
  • Not tested on real hospital notes. The Spanish PII test is held-out groups of the same corpus file the models learned from. The scope models were tested on held-out specialties of one annotated corpus, not on other institutions or other kinds of notes. Our note is synthetic.
  • Not validated as clinical PII models beyond Spanish. The other six languages were measured on synthetic finance documents, and German names stay below the earlier OpenMed model.
  • Not typed. The PII models return one generic label, so replacing names with fake names or shifting dates needs another step.
  • Not a full assertion model. The scope models read negation and uncertainty only. Outside a scope does not mean a condition is present, and they do not handle family history, time, hypotheticals, or who the condition belongs to. Uncertainty is their weak class.
  • Not probabilities. Scores rank spans and words; they are not calibrated.
  • Not medical devices. They find text spans and label words. They do not give medical advice.

Built on open work

These models stand on open work. Thank you to Fastino for GLiNER2.5 Multi, GLiNER2.5 Decide 1B and the GLiNER2 library, all Apache-2.0, and to Johns Hopkins University CLSP for mmBERT-base (MIT), the encoder behind two of the four. Thank you to Marimon and colleagues for the MEDDOCAN corpus (2019, CC BY 4.0), to Gretel for the synthetic PII finance multilingual dataset (Apache-2.0), to Leitner and colleagues for the German LER corpus (CC BY 4.0), and to the annotators of the Spanish clinical text the scope models learned from. No corpus text ships in any of the four repositories.

And the next time a search counts "tromboembolismo pulmonar", it can check what the note said about it first.

Community

Sign up or log in to comment