LabLens / tests /test_core.py
Imene-ee's picture
Upload 19 files
dab7db2 verified
Raw History Blame Contribute Delete
7.7 kB
"""Tests for the safety-critical, deterministic parts of LabLens."""
import pytest
from lablens.catalog import load_catalog, match_test_name
from lablens.explain import explain, render_markdown, violates_guardrails
from lablens.extract import extract, parse_line, parse_range
from lablens.flagger import flag_all, flag_result
from lablens.models import Flag, Result
def R(key, value, unit, lo=None, hi=None):
return Result(test_key=key, raw_name=key, value=value, unit=unit, ref_low=lo, ref_high=hi)
# ---------- flagging ----------
def test_report_range_decides_low_high():
assert flag_result(R("hemoglobin", 11.0, "g/dL", 12, 16)).flag == Flag.LOW
assert flag_result(R("hemoglobin", 16.5, "g/dL", 12, 16)).flag == Flag.HIGH
assert flag_result(R("hemoglobin", 14.0, "g/dL", 12, 16)).flag == Flag.NORMAL
def test_critical_beats_range_and_converts_units():
# 6.8 mmol/L potassium is critical regardless of the printed range
assert flag_result(R("potassium", 6.8, "mmol/L", 3.5, 5.1)).flag == Flag.CRITICAL_HIGH
# French g/L glucose: 0.45 g/L = 45 mg/dL, below the 50 mg/dL critical threshold
assert flag_result(R("glucose", 0.45, "g/L", 0.70, 1.10)).flag == Flag.CRITICAL_LOW
# 4.5 g/L glucose = 450 mg/dL -> critical high
assert flag_result(R("glucose", 4.5, "g/L", 0.70, 1.10)).flag == Flag.CRITICAL_HIGH
def test_cell_count_units_giga_per_litre():
# 40 G/L platelets = 40 x10^3/µL -> critical low
assert flag_result(R("platelets", 40, "G/L", 150, 400)).flag == Flag.CRITICAL_LOW
def test_implausible_value_is_never_normal_or_urgent():
# 436% HbA1c is a dropped decimal point, not a medical emergency or a normal value
r = flag_result(R("hba1c", 436, "%", 4.0, 5.6))
assert r.flag == Flag.CHECK
a = flag_all([R("hemoglobin", 151, "g/dL", 12, 17)])
assert a.results[0].flag == Flag.CHECK and not a.urgent
def test_catalog_range_used_when_report_has_none():
r = flag_result(R("sodium", 150, "mmol/L"))
assert r.flag == Flag.HIGH and r.ref_source == "catalog"
def test_unknown_unit_skips_critical_but_uses_printed_range():
r = flag_result(R("potassium", 7.0, "zz/qq", 3.5, 5.1))
assert r.flag == Flag.HIGH # not escalated: we can't be sure of the unit
def test_unknown_unit_and_no_range_is_unknown():
assert flag_result(R("potassium", 4.0, "zz/qq")).flag == Flag.UNKNOWN
def test_urgent_flag_set():
a = flag_all([R("hemoglobin", 14, "g/dL", 12, 16), R("sodium", 118, "mmol/L", 135, 145)])
assert a.urgent
def test_every_critical_threshold_is_plausible():
for td in load_catalog().values():
for v in td.critical.values():
assert td.plausible[0] < v < td.plausible[1], td.key
# ---------- name matching ----------
@pytest.mark.parametrize("name,key", [
("Glycémie à jeun", "glucose"), ("GLYCEMIE A JEUN", "glucose"), ("Kaliémie", "potassium"),
("Hemiogiouine", "hemoglobin"), ("HBAIC", "hba1c"), ("Ldi-choiestéroi", "ldl"),
("Hdl-chatesteroi", "hdl"), ("K", "potassium"), ("NA", "sodium"), ("الكرياتينين", "creatinine"),
])
def test_name_matching(name, key):
assert match_test_name(name)[0] == key
@pytest.mark.parametrize("name", ["Patient", "Collection date", "No", "Dossier", "Page"])
def test_non_tests_do_not_match(name):
assert match_test_name(name)[0] is None
# ---------- line and range parsing ----------
@pytest.mark.parametrize("text,expected", [
("12.0 - 16.0", (12.0, 16.0)), ("(0,70-1,10)", (0.70, 1.10)), ("[4,00 – 5,60]", (4.0, 5.6)),
("< 2,00", (None, 2.0)), ("> 0,40", (0.40, None)), ("[4.20 - 5.901", (4.20, 5.90)), ("", (None, None)),
])
def test_parse_range(text, expected):
assert parse_range(text) == expected
def test_parse_line_french_with_marker():
r = parse_line("Kaliémie 5,36H mmol/L 3,50 - 5,10")
assert (r.test_key, r.value, r.unit, r.ref_low, r.ref_high) == ("potassium", 5.36, "mmol/L", 3.5, 5.1)
def test_parse_line_skips_ocr_noise_and_stray_digit():
assert parse_line("TSH ae 3.11 mIU/L [0.40 - 4.00]").value == 3.11
assert parse_line("K 7 3.79 mmol/L [3.50 - 5.10]").value == 3.79
def test_parse_line_ignores_headers():
assert parse_line("Date de prélèvement: 2026-04-05") is None
assert parse_line("Patient: Karim Mansour") is None
# ---------- end to end on samples ----------
def test_sample_pdf_end_to_end():
results, _, lang, method = extract("samples/sample_en_report.pdf")
assert method == "pdf-text" and lang == "en" and len(results) == 13
def test_sample_urgent_photo_end_to_end():
results, _, lang, method = extract("samples/sample_fr_urgent.png")
a = flag_all(results)
assert method == "ocr" and lang == "fr" and a.urgent
# ---------- explanation guardrails ----------
@pytest.mark.parametrize("text", [
"You have diabetes.", "Vous avez une anémie.", "Take 1000 mg of iron daily.", "Stop taking your pills.",
])
def test_guardrails_block_diagnosis_and_dosing(text):
assert violates_guardrails(text)
def test_guardrails_allow_hedged_language():
assert not violates_guardrails("A high value can be linked to diabetes; your doctor can confirm.")
def test_template_explanation_without_api_key(monkeypatch):
monkeypatch.delenv("HF_TOKEN", raising=False)
monkeypatch.delenv("LLM_API_KEY", raising=False)
a = flag_all([R("ferritin", 8, "ng/mL", 15, 200), R("tsh", 2.0, "mIU/L", 0.4, 4.0)])
exp, source = explain(a, "fr")
assert source.startswith("template")
md = render_markdown(a, exp, "fr")
assert "Ferritine" in md and "Bas" in md and len(exp["questions"]) == 1
# ---------- LLM path (mocked: no network in tests) ----------
class _FakeClient:
def __init__(self, content):
msg = type("M", (), {"content": content})
choice = type("C", (), {"message": msg})
self.chat = type("Ch", (), {"completions": type("Co", (), {
"create": staticmethod(lambda **kw: type("R", (), {"choices": [choice]}))})})
def _patch_llm(monkeypatch, content):
import openai
monkeypatch.setenv("HF_TOKEN", "test")
monkeypatch.setattr(openai, "OpenAI", lambda **kw: _FakeClient(content))
def test_llm_output_used_when_safe(monkeypatch):
_patch_llm(monkeypatch, '{"summary": "Mostly fine.", "results": [{"key": "ferritin", "explanation": '
'"Low iron stores."}, {"key": "made_up", "explanation": "x"}], "questions": ["Why?"]}')
a = flag_all([R("ferritin", 8, "ng/mL", 15, 200)])
exp, source = explain(a, "en")
assert source == "llm"
assert [x["key"] for x in exp["results"]] == ["ferritin"] # hallucinated test dropped
def test_llm_diagnosis_falls_back_to_template(monkeypatch):
_patch_llm(monkeypatch, '{"summary": "You have anemia.", "results": [], "questions": []}')
a = flag_all([R("hemoglobin", 9, "g/dL", 12, 16)])
exp, source = explain(a, "en")
assert source.startswith("template") and "guardrail" in source
def test_llm_garbage_falls_back_to_template(monkeypatch):
_patch_llm(monkeypatch, "Sorry, I can't help with that.")
exp, source = explain(flag_all([R("tsh", 9, "mIU/L", 0.4, 4.0)]), "ar")
assert source.startswith("template")
def test_lab_marker_contradiction_becomes_check():
r = R("ast", 6.5, "UI/L", 3.0, 48.0) # range misread (8,0 -> 3,0) but lab printed "L"
r.report_marker = "L"
assert flag_result(r).flag == Flag.CHECK
ok = R("ast", 6.5, "UI/L", 8.0, 48.0)
ok.report_marker = "L"
assert flag_result(ok).flag == Flag.LOW
def test_glued_marker_with_trailing_punctuation():
r = parse_line("Vitamine d 22L. ng/mL 30 - 100")
assert (r.value, r.report_marker, r.ref_low) == (22.0, "L", 30.0)