DataStoryMD · defensible findings from messy clinical data

From the spreadsheet you actually have

Real-world data is a mess. That is the starting point, not the excuse.

Coded columns, value legends jammed into headers, sentinels, mixed languages. DataStoryMD maps every column to clinical meaning before it models anything.

Data Table · 300 rows · 133 columns 126 mapped7 unmapped

✓Age	✓Ethnicity	✓Family History	✓Diabetes	✓Debulking
67	Ashkenazi Jewish	No	No Diabetes	R0 (complete)
60	Non-Ashkenazi	Yes	No Diabetes	R2
41	Ashkenazi Jewish	No	Diabetes	R0 (complete)
72	Unknown	No	No Diabetes	R1
55	Ashkenazi Jewish	Yes	Diabetes	No surgery

BRCA1 carrier → Platinum sensitivityendpoint

OR 6.2 (95% CI 2.4–16.1) · q < 0.001 · n = 300

Named, correctly typed, modeled, and written up in a sentence a clinician can read aloud.

Beyond the findings

Ask it questions. Export the answers. Bring your own specialty.

The parts a findings list can't give you.

Ask in plain English

Type a clinical question and get a real answer with the statistics behind it. No query language, no pivot tables.

"Does complete debulking help platinum-resistant patients?"

Export for grants and talks

One click to a formatted Word or PowerPoint document, figures, methods, and citations included.

It already knows your field

Each specialty pack encodes the causal structure, temporal order, confounders, valid endpoints, and the traps. The statistical judgment is built in, so your clinical expertise is all it takes to run rigorous research.

Findings, organized

Not a flat list. Findings grouped into themes, ranked by impact.

Dozens of related results collapse into a handful of clinical themes, each scored and ready to open.

154comparisons tested → 154eligible to model → 65significant, FDR-corrected → 88stories · 89 not significant

Counts from the bundled 300-patient synthetic cohort. TCGA-OV, with 28 mapped clinical fields, tests 19.

Predictors of Overall Survival

6 converging stories

6 exploratory

Overall Survival differs by BRCA2 Carrier

BRCA2 Carrier

IMPACT97

BRCA / HRD group: 36.4-month gap in OS

BRCA / HRD Group

IMPACT60

Overall Survival differs by HRD Score

HRD Score

IMPACT40

Predictors of Platinum Sensitivity

4 converging stories

4 exploratory

BRCA1 Carrier: higher Platinum Sensitivity

BRCA1 Carrier · OR 6.2

IMPACT96

LOF Variant: higher Platinum Sensitivity

Loss-of-Function · OR 4.1

IMPACT96

Founder Mutation: higher sensitivity

Founder Mutation · OR 3.9

IMPACT44

Then keep going: the pathway map, cohort comparison, and live subgroup drill-down are built to be explored, not screenshotted. See them in the live demo →

Rigor that earns trust

It separates real effects from artifacts of study design.

Any tool can surface a correlation. The harder question is which associations reflect biology and which are artifacts of how the data was collected.

On this cohort DataStoryMD set aside 77 outcome candidates as ineligible and excluded hundreds of biased comparisons, each with a stated reason.

Excludednumber of treatment lines

Immortal-time bias

More treatment lines means the patient lived long enough to receive them. Survival drives the count, not the reverse.

No black box

Every number is one you could defend.

Adjusted models, honest reporting, and your data staying with you are the defaults, not add-ons. (This demo runs on a synthetic cohort with planted effects, so you can watch the tool recover exactly what was built in.)

300

synthetic patients, 133 raw columns

findings surfaced in one run

ineligible outcomes correctly excluded

Adjusted by default. Every effect comes from a multivariable model that adjusts for the confounders your domain pack pre-specifies, never a raw two-way test.
Reproducible, not generated. The same cohort always yields the same findings, from a fixed statistical engine, not a language model. Every number traces back to the rows and the test behind it.
Nothing invented. When an association is not significant, the tool says so plainly rather than dressing it up.
Your data stays with you. This public demo has no upload. Run your own cohort locally.

Checked against numbers we did not produce

A synthetic cohort proves it finds what was planted. Published research proves it is right.

So we ran it on cohorts whose results are already in the literature and compared every number we could check — against a standard textbook, a peer-reviewed paper, and the consortium that assembled the data.

published numbers checked

100%

textbook coefficients matched within 0.1%

3 of 3

published hazard ratios reproduced, each inside the other's interval

Ovarian cancer · TCGA, 316 patients	Published	DataStoryMD
Age at diagnosis → progression-free survival	0.995 (0.982–1.009)	0.995 (0.982–1.008)	matches
Age at diagnosis → overall survival	1.019 (1.005–1.033)	1.017 (1.004–1.031)	matches
Tumour stage → overall survival	1.325 (0.960–1.828)	1.267 (0.893–1.797)	matches

Hazard ratios from Birkbak et al., PLoS One 2013, reproduced on the same 316 patients that paper analysed. Every published estimate falls inside the interval DataStoryMD produced — and the stage row reproduces its non-significance too, not just its point estimate. Three further terms match; the full comparison, including one we do not match and why, is in the validation report.

A textbook, coefficient by coefficient. 77 published values from Hosmer & Lemeshow’s Applied Logistic Regression across eight model specifications — all within 0.1%. The three largest deviations are the book’s own rounding: it prints 4.87, we compute 4.867397.
The consortium’s own description of its cohort. All five figures the TCGA Research Network reports for these 489 patients reproduce — cohort size, survival status, platinum resistance and median follow-up.
Corrections these analyses did not apply. Multiplicity correction across every tested pair, proportional-hazards testing on every survival model, a confidence interval on every median — and a median reported only when the curve reaches it. On this cohort two associations reaching p<0.05 do not survive correction, and we report them as non-significant.
Where we differ, we say why. One published estimate we do not reproduce closely, because that paper cuts residual disease at different boundaries than our source file records. We publish the gap rather than tune it away.

Open the TCGA-OV cohort → Open the synthetic cohort

Sources: Hosmer & Lemeshow, Applied Logistic Regression 2nd ed. · Birkbak et al., PLoS One 2013;8(11):e80023 · TCGA Research Network, Nature 2011;474:609–615. Based upon data generated by the TCGA Research Network.

The findings are
already in your data.