MedTEMPO-Bench

MedTEMPO-Bench is a longitudinal medication-reconciliation benchmark across hospital stays. When a patient is readmitted, the prior record holds discharge medications, later outpatient and pharmacy notes, labs and orders, and part of it has already been overturned by the time of admission: a drug was stopped, switched or re-dosed, or a newer result replaced an older one. The agent plays the admitting physician and has to tell which records still hold at this moment, not merely whether it read them.

⚠️ Research use only. All patients and records are synthetic. This dataset is for benchmarking AI agents and must not be used for diagnosis or treatment decisions.

News

• 2026-10-30 — First public leaderboard dataset: synth_core200_v2n-20261028 (batch 202610, 1,000 scoring units = 200 items × 5 metrics). Item content and per-unit baseline results are public; the ground truth is withheld while this batch is the active leaderboard set (see Ground Truth).

Leaderboard

Scores are computed and published on HMA (Health Memory Arena):

• Leaderboard: https://healthmemoryarena.ai/leaderboard?bench=medtempo
• Dataset page: https://healthmemoryarena.ai/dataset

Dataset Summary

Item	Value
Items	200 (one per synthetic patient)
Admission index k	k ≥ 2; five strata of 40 items: 2nd / 3rd / 4–5th / 6–9th / ≥10th stay (up to the 20th)
Scoring units	1,000 (one per item × metric)
Visible evidence	154,730 pieces (labs 68,846 · orders 34,410 · discharge-summary sections 28,473 · outpatient and pharmacy notes 23,001); median 475 per item, P90 1,789
Ground truth	Ingredient-level (RxNorm) medication state at admission: continue / stop / new; 2,206 entries (continue 58.2%, stop 34.3%, new 7.6%); every item contains at least one stop
Data	Fully synthetic; no real patient content

Task

One item = one synthetic patient at the moment t_k of their k-th admission (k ≥ 2). The agent receives:

• age, sex and chief complaint (no history of present illness, which would restate the medication changes between stays);
• every piece of evidence from the previous k − 1 stays and the outpatient period in between whose record time is earlier than t_k — discharge-summary sections, discharge diagnoses, inpatient orders, labs, outpatient and pharmacy notes — each with an id, event time, record time and source.

It submits a structured answer: the current medication list (each ingredient marked continue / stop / new), a problem list and a principal diagnosis, with every claim supported by cited evidence ids.

Metrics

Five metrics per item, each in 0–1 (shown × 100), higher is better:

Metric	Tag	Definition
Medication F1	metric:med_f1	Ingredient-level F1 of the medication list against the truth; a hit needs the right ingredient and the right continue / stop / new call
Stale-evidence avoidance	metric:semr	1 − SEMR, where SEMR is the share of citations supporting a "still holds" claim that point at evidence already stale at t_k
Over-stopping avoidance	metric:overstop	1 − the share of drugs called "stop" that are in fact still taken
Stop recall	metric:stop_recall	Share of truly stopped drugs that were called "stop"
Valid-answer rate	metric:failure	1 if the item produced a valid structured answer, else 0

Read the middle three together: a system that stops everything looks clean on stale-evidence avoidance, and a system that never stops looks clean on over-stopping avoidance; stop recall exposes both.

Dataset Structure

data/202610/
├── synth_core200_v2n-20261028.jsonl          # scoring-unit index (1,000 rows)
├── synth_core200_v2n/
│   ├── MANIFEST.json                         # bank fingerprint + sha256 of every bank file (truth files included)
│   ├── cohort_synth_core200_v2n.json         # the 200 (subject_id, hadm_id) items
│   ├── units/synth_core200_v2n/<item>.json   # visible evidence for each item
│   └── presentation/synth_core200_v2n/<item>.json  # age, sex, chief complaint
└── reports/<dataset>_medtempo_<agent>.json   # per-unit baseline results (9 files)
manifest.json                                 # change-detection entry point

<item> is <subject_id>_<hadm_id>, e.g. 90200027_9020002708.

Scoring-unit index (synth_core200_v2n-20261028.jsonl)

{"id": "90200027_9020002708__med_f1", "unit": "90200027_9020002708", "metric": "med_f1", "tags": ["metric:med_f1"]}

Each row holds only the unit id, the item, the metric and its tag — no evidence and no answers.

presentation/<item>.json

{"subject_id": "90200027", "hadm_id": "9020002708", "chief_complaint": "Dysuria, fever", "age": 46, "sex": "M", "hpi": null, "scrubbed": []}

units/<item>.json

Field	Description
subject_id, hadm_id	Synthetic patient and admission ids
k	Admission index of this item
t_k	Admission time; only evidence recorded before it is visible
sources	Evidence sources present for this item
visible[]	The evidence list (fields below)

Each visible[] entry:

{"unit_id": "U90200027_rx1013", "admission_id": "9020002701", "text": "医嘱：Cefepime 1 g IV（2144-09-11 至 2144-09-13）", "event_time": "2144-09-11T21:44:51", "record_time": "2144-09-11T21:56:35", "time_source": "native", "source_tier": "order", "concept": "RxNorm:20481"}

Field	Description
unit_id	Evidence id the agent cites
admission_id	The earlier stay this evidence belongs to
text	Rendered evidence text (clinical content in English; short labels such as 医嘱： "order", 门诊记录 "outpatient note" and 参考 "reference range" are Chinese)
event_time / record_time	When it happened / when it was written; visibility is decided by record_time
time_source	native or inferred
source_tier	discharge_summary, physician_note, order, lab, …
concept	Code of the underlying concept (RxNorm, LOINC, ICD-10-CM)

Dates are shifted into fictional years (21xx).

reports/<dataset>_medtempo_<agent>.json

One file per leaderboard row. target names the agent, model and arm (reference: true marks the shortcut row); summary.dimension_scores holds the five metric scores; summary.failed_units lists items that failed to produce a valid answer; cases[] has one entry per scoring unit:

{"id": "90200027_9020002708__med_f1", "tags": ["metric:med_f1"], "eval": {"score": 0.8462, "result": "scored"}}

eval.score is null when the metric does not apply to that item (zero denominator) or cannot be computed because the item failed.

Construction

Items come out of a generative process in two layers:

• Fact layer (no LLM): cohort and timeline → latent physiological state → clinical decisions (discharge medications, stops / dose changes / new drugs between stays, orders, outpatient and external labs, diagnoses). Every patient's randomness is reproducible from a seed.
• Record layer: structured sections (medication lists, diagnoses, orders, labs, outpatient lines) are rendered by code; only narrative sections are written by a model (gemini-3.7-flash for 101 items, mistral-medium-3.1 for 99, assigned at random) from the facts, and discharge summaries are accepted only after a four-dimension review by claude-sonnet-5. Record-layer noise is injected (external pharmacy snapshots, outpatient corrections of drug names, inpatient formulary substitutions); 180 items show it.

Generation parameters were calibrated against aggregate statistics only (counts, with small cells below 100 suppressed); no real patient record or text enters the generator.

Item selection is independent of any evaluated model and of the answer mix: within each k-stratum the top 40 items are taken by a quality score built only from construction-time readings and clinical-review scores, and the rule was committed before the selection was seen. Readings directly tied to the answer mix (proximal stale drugs, change events in the last interval, single-item shortcut scores) carry no weight.

Ground Truth

The truth is derived directly from the fact layer, with no model involved, and the evaluated system's inferences never enter any label:

• the validity of every visible piece of evidence at t_k (valid 31.0% · superseded 23.4% · expired 18.9% · unknowable 26.8%);
• the ingredient-level medication state at the k-th admission (continue / stop / new);
• the principal diagnosis (ICD-10).

"New" only counts ingredients that appear in visible evidence after the last discharge and before t_k (evaluation protocol eval-protocol-v2); untraceable new drugs are removed from the truth, and submissions on them are not scored.

The truth files are not in this repository while batch 202610 is the active leaderboard set. MANIFEST.json already carries their sha256, so the files can be verified byte for byte when they are released after the batch is retired.

Scoring

• All scores are computed by code; there is no LLM judge.
• An item that fails to run (no valid answer, truncated output or content filter) scores 0 on medication F1, with all items in the denominator; the other three quality metrics cannot be computed on it and are left out, which the valid-answer rate column makes visible.
• A metric that does not apply to an item (zero denominator, e.g. no drug called "stop" means no over-stopping rate) is left out of that metric's denominator — "not applicable" is not a full score.
• Metric score = mean over scored units × 100. There is no total score.

Known Limitations

• Stops are over-represented: 34.3% of truth entries, against about 21% in the aggregate real-world profile (synthetic patients have denser change events between stays).
• The zero-reasoning shortcut scores 57.1 on medication F1 because the truth is mostly "continue"; read each model's gap to the reference row rather than absolute scores.
• "New" carries little weight (7.6%; 102 items have no new drug), so medication F1 mostly tests continue versus stop.
• How outpatient medication evidence is worded is a synthetic assumption.
• No clinician has reviewed the items; clinical plausibility was checked by an LLM reviewer only.
• One sample per item and model: readings carry patient-sampling variance only, not rerun variance.

manifest.json Format

{
  "version": 1,
  "updated_at": "…",
  "latest": "data/202610/synth_core200_v2n-20261028.jsonl",
  "batches": {
    "202610": {
      "created_at": "…",
      "checksum": "sha256:…",
      "user_count": 0,
      "eval_benchmark": "medtempo",
      "eval_dataset": "synth_core200_v2n-20261028",
      "files": ["data/202610/synth_core200_v2n-20261028.jsonl", "…"]
    }
  }
}

Field	Description
version	Increments by 1 on every change
latest	Path of the current leaderboard dataset, data/<batch>/<eval_dataset>.jsonl
batches.<batch>.checksum	SHA-256 over all files of the batch
batches.<batch>.user_count	Always 0 (MedTEMPO-Bench has no virtual users)
batches.<batch>.eval_benchmark	Always medtempo
batches.<batch>.eval_dataset	Dataset name; the index file is data/<batch>/<eval_dataset>.jsonl
batches.<batch>.files	Every file of the batch

Batches are named YYYYMM. Earlier batches stay in the repository when a new one is added.

Usage

import json
from pathlib import Path
from huggingface_hub import snapshot_download

root = Path(snapshot_download(repo_id="mirobody/MedTEMPO-Bench", repo_type="dataset"))
bank = root / "data/202610/synth_core200_v2n"
item = "90200027_9020002708"
unit = json.loads((bank / f"units/synth_core200_v2n/{item}.json").read_text())
pres = json.loads((bank / f"presentation/synth_core200_v2n/{item}.json").read_text())
print(pres["chief_complaint"], unit["k"], unit["t_k"], len(unit["visible"]))

from datasets import load_dataset

units = load_dataset("mirobody/MedTEMPO-Bench", "synth_core200_v2n-20261028", split="test")

License

Released under CC BY 4.0. Evidence carries RxNorm, LOINC and ICD-10-CM codes; use of those vocabularies is subject to their own terms.

Citation

@misc{medtempobench2026,
  title        = {MedTEMPO-Bench: Longitudinal Medication Reconciliation Across Hospital Stays},
  author       = {Ang Gao and Cailiang Liu},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/mirobody/MedTEMPO-Bench}}
}