Skip to content

How to Interpret Your Own Lab Results

Oak
Glowing reed stalks in a dark marsh, each with a teal band and a few amber nodules above and below it.

By the end of this guide you will have every blood test you have ever taken in a single tidy table. Each row is one result carrying a LOINC code and a UCUM unit. It also carries a collection timestamp, the performing lab, and the reference interval that lab used. On top of that table you will have three derived quantities that matter more than the high/low flag on the PDF: a within-person Z-score, a reference change value that tells you whether a change between two draws is larger than measurement noise, and a lab-to-lab bias annotation so you do not mistake a vendor switch for a physiological event. You need your results in electronic form, Python 3.11 with pandas, and about an hour. The electronic form can be a FHIR API export, an HL7 v2 message dump, or a CSV from your portal. Everything here is measurement and interpretation. It is not diagnosis, and several steps end with the recommendation to bring a specific number to a clinician.

1. Pull the results in a structured format and not as PDFs

Most US portals expose a patient-facing FHIR R4 API under the 21st Century Cures Act rules, which is the format you want because each result arrives as an Observation resource with a code, a value, a unit, and a timestamp already separated. A typical pull, once you have an access token from the portal’s SMART-on-FHIR app registration:

curl -H "Authorization: Bearer $TOKEN" \
     -H "Accept: application/fhir+json" \
     "$BASE/Observation?patient=$PATIENT_ID&category=laboratory&_count=200" \
  > obs_page1.json

Follow the link[relation=next] URL until it disappears. If your lab only offers a CSV or a CCDA XML download, that is workable, but expect to do the coding step by hand. Scraping PDFs is the last resort. If you must, pdfplumber with extract_table() beats OCR. Hand-check every row, because a decimal misread in a creatinine value is invisible downstream.

Flatten the FHIR bundle into rows with jq before you touch pandas:

jq -r '.entry[].resource
  | select(.resourceType=="Observation")
  | [ (.effectiveDateTime // .effectivePeriod.start),
      (.code.coding[] | select(.system|test("loinc")) | .code),
      .code.text,
      (.valueQuantity.value // ""),
      (.valueQuantity.unit // ""),
      (.referenceRange[0].low.value // ""),
      (.referenceRange[0].high.value // ""),
      (.performer[0].display // "") ]
  | @csv' obs_page*.json > labs_raw.csv

Two things go wrong here reliably. Panels arrive as a parent Observation with hasMember pointers and no value of their own, so you will see empty rows for “Comprehensive metabolic panel” that you should drop. And some vendors put the LOINC code only on the panel, coding the children with a local mnemonic, in which case you map codes yourself in step 2.

2. Decode the codes and the abbreviations

The codes on a blood test come in two layers. The local layer is the lab’s internal mnemonic, which is not portable. Examples are NA, CREA, ALT, HGB, and WBC. The portable layer is LOINC, a six-or-seven digit identifier with a check digit, maintained by the Regenstrief Institute. A LOINC code specifies six axes: the component measured, the property (mass concentration versus substance concentration versus catalytic activity), the timing, the specimen, the scale, and where relevant the method. That is why total cholesterol in serum reported as mass per volume is 2093-3, and why a different code exists for the same analyte in a different specimen or units.

Codes you will meet constantly are the following.

  • 718-7 hemoglobin in blood
  • 6690-2 leukocytes
  • 777-3 platelets
  • 2951-2 sodium in serum or plasma
  • 2160-0 creatinine
  • 1742-6 alanine aminotransferase
  • 1920-8 aspartate aminotransferase
  • 2093-3 total cholesterol
  • 2085-9 HDL cholesterol
  • 2571-8 triglycerides
  • 13457-7 LDL cholesterol calculated by Friedewald
  • 4548-4 hemoglobin A1c
  • 3016-3 thyrotropin
  • 30522-7 high-sensitivity C-reactive protein

Download the full LOINC table from Regenstrief and join on it rather than hardcoding a dictionary. It is free but requires registration. The description text tells you the property and specimen, which is exactly the information that prevents you from plotting mg/dL and mmol/L on the same axis.

The panel abbreviations themselves decompose predictably. A CBC (complete blood count) is the cellular compartment.

  • WBC with a five-part differential (neutrophils, lymphocytes, monocytes, eosinophils, and basophils)
  • RBC with its indices MCV, MCH, MCHC, and RDW
  • hemoglobin
  • hematocrit
  • platelets with MPV

A CMP (comprehensive metabolic panel) is the following.

  • electrolytes (Na, K, Cl, and CO2 as bicarbonate)
  • renal markers (BUN, creatinine, eGFR)
  • glucose
  • calcium
  • total protein and albumin
  • liver enzymes (ALT, AST, ALP, and total bilirubin)

A lipid panel is total cholesterol, HDL-C, triglycerides, and an LDL-C that is usually calculated rather than measured. Everything else is an add-on.

3. Normalize units before you do any arithmetic

Unit chaos is the single largest source of self-inflicted error in personal lab analysis, and it is worst in exactly the analytes people track most. Map every unit string to UCUM, then convert to one canonical unit per LOINC code.

import pandas as pd

TO_SI = {                     # (loinc, from_unit) -> (factor, to_unit)
    ("2093-3", "mg/dL"): (0.02586, "mmol/L"),   # total cholesterol
    ("2085-9", "mg/dL"): (0.02586, "mmol/L"),   # HDL-C
    ("13457-7","mg/dL"): (0.02586, "mmol/L"),   # LDL-C
    ("2571-8", "mg/dL"): (0.01129, "mmol/L"),   # triglycerides
    ("2345-7", "mg/dL"): (0.05551, "mmol/L"),   # glucose
    ("2160-0", "mg/dL"): (88.4,    "umol/L"),   # creatinine
}

def to_si(row):
    key = (row.loinc, row.unit)
    if key in TO_SI:
        f, u = TO_SI[key]
        return pd.Series([row.value * f, u])
    return pd.Series([row.value, row.unit])

df[["value_si", "unit_si"]] = df.apply(to_si, axis=1)

Keep both columns. You will want mg/dL when reading American reference ranges and mmol/L when reading European papers. Hemoglobin A1c has its own trap: NGSP percent and IFCC mmol/mol are related by IFCC = (NGSP − 2.15) × 10.929. A value of “48” is an IFCC result. It is not a catastrophic percentage.

4. Treat the reference interval as a population statistic and not a verdict

A reference interval is almost always the central 95% of results from a reference population the lab selected, which means 1 in 20 healthy people fall outside it for any given analyte by construction. On a 20-analyte panel, if the analytes were independent, the probability of at least one flagged value is 1 − 0.95²⁰ ≈ 0.64. They are not independent, so the real number is lower, but the intuition holds: a single asterisk on a wide panel carries very little information on its own.

Reference intervals are also partitioned imperfectly. Many are partitioned by sex and some by age, and there is good evidence that routine hematology and chemistry markers carry sex-specific structure strong enough that machine learning models separate phenotypes differently in men and women 1. Population studies that collect venous blood at scale also show how much the interval depends on the assay platform, the fasting state, and the collection and processing protocol. That is why a well-documented study specifies all three before reporting distributions 2. Two practical consequences: compare your result to the interval printed by the lab that ran it, and never compare a result to an interval you found on the internet for a different assay.

5. Quantify the noise before you interpret a change

The question that matters for longitudinal data is not whether a value is inside the interval but whether a change between two draws is bigger than the combined analytical and biological variation. The standard tool is the reference change value:

RCV = 2^(1/2) × Z × sqrt(CV_A² + CV_I²)

where CV_A is the analytical coefficient of variation for the assay (ask your lab, or read it off their method sheet) and CV_I is the within-subject biological coefficient of variation, available from the EFLM Biological Variation Database. Z is 1.96 for a two-sided 95% criterion, 1.65 for one-sided 95%.

import math

def rcv(cv_a, cv_i, z=1.96):
    return math.sqrt(2) * z * math.sqrt(cv_a**2 + cv_i**2)

rcv(2.0, 5.9)    # creatinine-like: ~17% change needed
rcv(3.0, 19.0)   # ALT-like: ~53% change needed

The numbers are sobering. For a tightly regulated analyte such as sodium, where CV_I is well under 1%, a 3 mmol/L move is large. For ALT or TSH, where CV_I is around 15 to 20%, a result that doubles between two draws may sit inside ordinary fluctuation. Compute the RCV once per analyte, store it in your table, and only look at deltas that exceed it. This is also the to-the-data answer to “my number went up, should I worry”. If the change is under the RCV, the measurement cannot distinguish it from noise, and the next step is repeating the draw rather than reasoning about it.

Analytical validity deserves the same scrutiny you would give any measurement instrument. Reviews of biomarker validation consistently separate analytical performance from validity as a marker of the biology you care about. Analytical performance covers precision, accuracy, matrix effects, and stability. A marker can pass the first and fail the second 3.

6. Build within-person Z-scores and personal ranges

Once you have four or more results for an analyte collected under comparable conditions, your own distribution is more informative than the population’s. Compute a within-person Z using a median and a median absolute deviation so that one outlier does not inflate the spread.

import numpy as np

def personal_z(s):
    med = s.median()
    mad = (s - med).abs().median() * 1.4826    # ~sigma for normal data
    return (s - med) / mad if mad > 0 else s * 0

df["z_self"] = df.groupby("loinc")["value_si"].transform(personal_z)

The relevant statistic for deciding whether this is worth doing is the index of individuality, CV_I divided by CV_G (between-subject variation). When it falls below roughly 0.6, the population interval is a poor detector for the individual: your value can travel most of the way across the interval while remaining far from your own set point. Creatinine, TSH, and many hormones behave this way. That is the statistical argument for repeated measurement of the same person over time rather than a single cross-sectional panel.

7. Annotate lab and platform changes as covariates

Different manufacturers’ assays for the same analyte are calibrated to different traceability chains, and the resulting bias is a step function in your time series. Add performer to every row, along with the device or method if available. Then check for a level shift whenever it changes. A crude but effective test is to fit a segmented model and compare it to a single-level model.

The lipid panel is the clearest example of why method matters. A calculated LDL-C is a derived quantity, not a measurement, and it diverges from directly measured values at high triglycerides and at low LDL-C. Large clinical laboratory databases that measure cholesterol synthesis and absorption markers alongside the standard panel show substantial heterogeneity in cholesterol homeostasis among people whose routine lipid numbers look similar 4. If your lipid panel is the thing you care about, apolipoprotein B and lipoprotein(a) are measured directly and are worth the extra order, and the interpretation of them belongs in a conversation with a clinician.

8. Read a single abnormal value in context, and know when it is a rule-out

Individual markers behave very differently depending on whether they are being used to rule in or rule out. Meta-analyses of blood biomarkers for excluding intracranial abnormality after mild head injury illustrate the pattern: the value of the test lies in a high negative predictive value at a chosen cutoff, and that NPV depends on the prevalence in the population being tested, so the same cutoff performs differently in different settings 5. When you read a cutoff in a paper, check the prevalence of the underlying condition in the study population before you apply the cutoff to yourself.

Acute-phase markers add another dimension: timing. CRP, procalcitonin, and related markers move on characteristic kinetics after a stimulus. Their perioperative interpretation depends heavily on how many days have elapsed since the insult 6. A CRP drawn two days after a hard training session, an infection, or a dental procedure is measuring that event. Record context alongside the value: fasting status, time of day, days since last intense exercise, illness, and menstrual cycle phase where relevant. Without those covariates, half your outliers are unexplainable.

Finally, note where the frontier is. Transcriptional signatures from whole blood can separate disease states that no single chemistry value distinguishes, as with the interferon-inducible, neutrophil-driven signature described in tuberculosis 7. Expression-level markers have also been proposed for time-sensitive processes such as post-stroke blood brain barrier disruption 8. These are research-grade readouts. They are informative for understanding your own biology over time, and they are not a basis for self-directed action.

Common problems

Duplicate rows from overlapping FHIR pages. Deduplicate on the tuple of (effectiveDateTime, loinc, value, performer) rather than on resource id, because some servers re-issue ids across pages.

Values reported as strings with comparators. <5, >1000, and NEGATIVE will break numeric parsing. Keep a comparator column, store the numeric part separately, and exclude censored values from trend fitting or handle them explicitly as left-censored.

Missing or inconsistent reference ranges. Some vendors omit referenceRange entirely, others attach an adult range to a pediatric result. Prefer the range printed on the original report from the performing lab, and record which lab it came from.

Local codes with no LOINC. Map them yourself using the lab’s method sheet and record the mapping in a version-controlled file with a note on specimen and units. Do not guess between two LOINC codes that differ only in specimen.

Timezone and date-only timestamps. Collection times that arrive as dates lose the diurnal signal. That matters for cortisol, iron, and testosterone. Where the time is missing, mark the row and exclude it from any diurnal analysis rather than imputing noon.

Over-reading a single flagged value. Before drawing any conclusion from an out-of-range result, check three things. Does the delta exceed the RCV, did the performing lab or method change, and does a pre-analytical covariate explain it? If it survives all three and concerns an organ system, take the specific number and its trend to a physician. A trend line plus the RCV computation is a much better input to that conversation than a PDF with an asterisk.

Footnotes

  1. Weizhe Zhen, Jingjing Chen, Hongjun Zhen, et al. Sex-related molecular phenotypes in anxiety-depressive disorders: a machine learning analysis of routine blood biomarkers. Frontiers in Psychiatry, 2026. https://doi.org/10.3389/fpsyt.2026.1889453 ↩

  2. Xinxin Chen, Eileen Crimmins, Peifeng (Perry) Hu, et al. Venous Blood-Based Biomarkers in the China Health and Retirement Longitudinal Study: Rationale, Design, and Results From the 2015 Wave. American Journal of Epidemiology, 2019. https://doi.org/10.1093/aje/kwz170 ↩

  3. Valentina Gallo, Aneire Khan, Carlos Gonzales, et al. Validation of biomarkers for the study of environmental carcinogens: a review. Biomarkers, 2008. https://doi.org/10.1080/13547500802054611 ↩

  4. Thomas D. Dayspring, Stephen A. Varvel, Leila Ghaedi, et al. Biomarkers of cholesterol homeostasis in a clinical laboratory database sample comprising 667,718 patients. Journal of Clinical Lipidology, 2015. https://doi.org/10.1016/j.jacl.2015.08.003 ↩

  5. Lara Milevoj Kopčinović, Chiara Di Resta, Sanja Stanković, et al. Diagnostic accuracy of blood biomarkers for excluding head CT abnormalities after mild traumatic brain injury in adults: a systematic review and meta-analysis. Critical Reviews in Clinical Laboratory Sciences, 2026. https://doi.org/10.1080/10408363.2026.2708643 ↩

  6. Simona Gigliotti, Giuseppe Guerriero, Giuseppe Mazza, et al. Perioperative Blood Biomarkers of Infectious and Non-Infectious Postoperative Pulmonary Complications: A Narrative Review. Journal of Clinical Medicine, 2026. https://doi.org/10.3390/jcm15020699 ↩

  7. Rafael Firszt, Brian Vickery. An Interferon-Inducible Neutrophil-Driven Blood Transcriptional Signature in Human Tuberculosis. Pediatrics, 2011. https://doi.org/10.1542/peds.2011-2107llll ↩

  8. Grant C. O’Connell, Madison B. Treadway, Ashley B. Petrone, et al. Peripheral blood AKAP7 expression as an early marker for lymphocyte-mediated post-stroke blood brain barrier disruption. Scientific Reports, 2017. https://doi.org/10.1038/s41598-017-01178-5 ↩

Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.