Find the trial site that stopped looking normal.
One site in forty is behaving differently. Which one?
Multi-site trials produce data of uneven quality. Sitewatch runs central statistical monitoring across every site every night and flags the ones whose data stopped resembling the rest, so monitoring visits go where the risk is. It flags sites for human review. It does not decide anything.
A monitoring signal only. It does not determine misconduct, does not unblind, and makes no clinical judgement.
Sixteen real series, one site diverging, detection computed live in your browser.
The problem
You cannot visit every site, and the sites you visit are chosen by rota.
On-site monitoring is the most expensive line in a trial budget and it is usually allocated by schedule rather than by risk. Meanwhile the signals that a site is drifting are already in the EDC: enrolment pace against the site’s own history, protocol deviation rate, query rate per form, and the spread of reported continuous measurements. Nobody watches those per site per day, so a site that changed in July gets noticed at a data review meeting in October, and by then the affected subject records are the ones the statistician has to decide what to do with.
The insight
Data that gets quieter while everything else gets louder.
Most monitoring rules look for values that are too high. The measure that actually separates a problem site from a busy site runs the other way: variability that is too low. Real biological measurements are noisy, and human beings inventing or rounding them are not, so a site whose reported continuous measurements tighten up is showing a signature that no threshold on that measurement would ever catch, because every individual value is perfectly plausible. Put that next to enrolment pace measured against the site’s own baseline rather than the study target, and you have a pattern worth a monitoring visit. Regulators already point at exactly this approach: central statistical monitoring is described in ICH E6(R2) and in FDA and EMA risk-based monitoring guidance as an acceptable alternative to visiting every site, and this is a working implementation of it rather than a policy document about it.
CUSUM on day-over-day percentage changes and Bayesian Online Changepoint Detection (Adams & MacKay 2007) run per site per measure, including a dispersion series so a variance collapse is detectable as a changepoint in its own right, with Benjamini-Hochberg FDR control across every site by measure combination so a forty-site study does not generate two referrals a week by arithmetic alone.
How it works
Four steps, no data science team
A standard extract from Medidata Rave, Veeva CDMS, OpenClinica, or a CDISC SDTM drop. Aggregate, de-identified, and blinded. No new site burden and nothing asked of investigators.
Enrolment pace against each site’s own baseline, protocol deviation rate, query rate per form, visit-window compliance, and the dispersion of reported continuous measurements, each held per site.
Every site by measure series goes through CUSUM and BOCD. What crosses is filtered by Benjamini-Hochberg across the whole study, because a forty-site study runs hundreds of tests a night and uncorrected testing at that width manufactures referrals.
Site, measures, date, and the statistical trace, routed into the risk-based monitoring cycle for a human to review. No conclusion about cause, no unblinding, no clinical judgement, no communication to the site that a person did not write.
Who it is for
The clinical operations lead who owns the monitoring plan
Clinical operations and data management at sponsors and CROs running multi-site interventional trials, typically twenty sites and up, already committed to a risk-based monitoring plan under ICH E6(R2) and looking for something to actually run it with.
Pricing
- –Nightly detection across every site
- –Full statistical trace per referral
- –Monitoring-plan export
- –Blinded, aggregate data only
- –Cross-study site history
- –Site risk profile that follows the site
- –EDC connectors
- –SSO, audit log, 21 CFR Part 11 support
- –Validation documentation package
- –VPC or on-premise deployment
- –Sponsor-level data segregation
- –Custom measures per protocol
- –Statistical review support
Competition
What exists, and what it does not do
| Who | What they do | The gap |
|---|---|---|
| CluePoints | Central statistical monitoring, the direct incumbent, used on large pharma programmes. | Sold and delivered as a programme with a services layer, priced for phase III at a large sponsor. Mid-size sponsors and CROs running twenty-site studies mostly do not buy it, so they run nothing at all. |
| Medidata Detect | Risk-based quality management inside the Rave platform. | Strongest if the study already lives entirely in Rave, and its risk indicators are largely threshold-based against study-wide expectations rather than changepoint-based against each site’s own history. |
| Veeva CDB | Cross-source clinical data aggregation and review. | Solves getting the data into one place, which is a real problem. It is a review surface, not a detector, so someone still has to look. |
| The KRI spreadsheet the CRO already maintains | Key risk indicators tracked monthly against fixed thresholds. | Fixed thresholds compare every site to a study-wide number, so a site that is anomalous relative to itself but inside the study band is invisible. It is also monthly, and it has no correction for the number of indicators it just checked. |
Two ways this dies. The regulatory one: this is a monitoring aid, and if a customer starts treating a referral as a finding about a site or an investigator, the product has caused harm and the liability is real. The mitigation is in the product surface itself, which states what it does not determine on every alert, but a mitigation in copy is weaker than a mitigation in process. The commercial one: clinical software sells on validation packages, quality agreements, and audit history, and a new vendor has none of those. Buying cycles run twelve to eighteen months and the first customer has to be willing to be the first customer in a domain where nobody wants to be.
Market
Priced against monitoring visit budget, which it partly replaces
A single on-site monitoring visit costs several thousand dollars once travel and monitor time are counted, and a sixty-site study schedules hundreds of them. Reallocating even a small share of those visits by risk pays for the Portfolio tier many times. There are a few thousand sponsors and CROs running multi-site interventional trials at any time; six hundred at the Portfolio tier is $65M ARR.