Feasibility work usually starts with how many people have the condition. Between that number and a randomised subject sit six or seven filters. Each sounds survivable on its own. Multiplied together, they can leave a pool far smaller than the recruitment plan assumed.
We build the analytics for deciding where to run a study and whether its patients exist: population funnel modelling, evidence-based site selection, criterion-level screen failure analysis, and early retention risk signals.
The characteristic failure in recruitment planning is not a single bad assumption. It is a chain of individually defensible ones. The condition affects a large population. Most of those patients are treated in settings where sites exist. A reasonable share will meet the criteria. Physicians will refer when the trial is relevant. Patients will usually consent when asked. Screening will pass most of those who consent. Read one at a time, none of those statements is obviously wrong. Multiplied, they are the difference between a plan that works and one that needs rescuing in month eight.
The second problem is that the data to do better mostly exists and is not assembled. Your completed studies hold screen failure patterns, consent rates, site-level conversion and discontinuation timing. Real-world sources can characterise how many patients in a geography carry the diagnosis and, with care, how many survive some criteria. Site databases hold what each site delivered before. None of this is exotic, and in most organizations it sits in separate places and gets replaced at feasibility time by a number somebody estimated in a meeting.
Feasibility counts prevalence and estimates the rest
When the filters after prevalence are where the number actually goes.
Screen failure is recorded as a coarse reason code
So nobody can say which criterion cost how many patients.
Site selection recommends the sites you already use
Because the performance data is about sites you have used.
Nobody counted the approved-contact stage
Potentially eligible and reachable through an approved process are different numbers.
Competing studies are not in the model
Your site is enrolling against three other protocols for the same patients.
Dropout is noticed at month six
When the behavioural signal was visible in the first two visits.
Referral volume is reported as recruitment performance
Referrals are not consents, consents are not randomisations.
Representativeness cannot be managed because it was not measured
And what was collected varies by region and is often optional.
Build the funnel backwards from a study you already completed. Take a finished study and reconstruct its funnel from the bottom: randomised subjects, screened, consented, approached, referred, identified as potentially eligible. Each ratio you can compute is a measured stage you no longer have to assume, and the shape of the losses is usually surprising. Most organizations discover that one stage they never measured accounts for the largest single loss, and that it is the same stage in every study they run. That is a reusable finding, it costs a week of work, and it makes the next feasibility exercise an arithmetic exercise rather than a negotiation.
The order places prediction after standardization for a reason that matters here more than on most pages. A site selection model trained on inconsistently defined historical performance will learn your recording conventions rather than your sites, and it will do so confidently.
Step 1
Assess the decision and its window
Step 2
Integrate site history, study history and population data
Step 3
Standardize site, investigator and funnel definitions
Step 4
Analyze the funnel and the losses at criterion level
Step 5
Predict with the survivorship problem acknowledged
Step 6
Visualize for the decision and its owner
Step 7
Monitor the funnel live, not at milestone review
Step 8
Optimize and feed findings into the next protocol
Ask what a site model would say about a site you have never used. Take your site selection approach and apply it to a plausible site with no history with your organization. If the method cannot produce a score, or produces one with no stated uncertainty, then the selection process can only re-rank the sites you already know. That is worth knowing explicitly, because it explains why site lists look similar study after study and why expansion into a new region is consistently harder than expected. The fix is a model built on site attributes rather than site identity, and it is ordinary work.
This page is the selection and funnel work. Managing a study that is already running, including milestones, startup queues, portfolio reporting and CRO oversight, is Clinical Operations Analytics. The two are deliberately separate because they are different decisions made by different people at different times, and they should be read together.
Decide where
Site feasibility analytics combining your study history with population and site data
Site selection intelligence scored on attributes, so sites without your history can be assessed
Investigator selection analytics including indication experience and research track record
Country and region prioritisation accounting for care pathway and competing trial load
Site mix scenario analysis, comparing candidate site lists on reach, speed and composition
Competing study visibility from registry data and feasibility responses recorded as data
Decide who
Patient population analysis from real-world sources within permitted use
Population funnel modelling with ranges and a stated source per stage
Eligibility criteria impact modelling against the structured criteria that can be estimated
Patient identification support for pre-screening, where consent and permitted use allow
Recruitment planning by channel, measured on randomisations rather than referrals
Measure the funnel
Recruitment funnel analytics across identified, approached, consented, screened and randomised
Enrollment forecasting at site and country level with ranges and recorded assumptions
Screen failure analysis at criterion level, with the failing value where it was captured
Conversion analytics by site, channel, referral source and geography
Screening workload and site capacity indicators, since a site can only screen so many patients a week
Diversity and representativeness measurement against the treated population, where collected
Keep the patients
Retention analytics including discontinuation timing, reason and site pattern
Dropout risk indicators built on early behavioural signals rather than late outcomes
Visit burden and schedule analysis, including travel and time cost where known
eCOA and ePRO compliance decay as an early retention signal
Patient support service utilisation measured against retention outcome
Site performance benchmarking on conversion and retention rather than raw enrollment
Site selection and protocol eligibility are close to irreversible. Once a site is contracted and a protocol is approved, the recruitment plan is largely determined and everything afterwards is management. That asymmetry is the argument for spending analytical effort before the study starts rather than reporting on it once it has.
| Context | The decision | What the analytics has to get right |
|---|---|---|
| Phase I | Which unit, and whether healthy volunteers or patients are reachable | Small numbers, so unit capacity and panel availability dominate |
| Phase II | Which sites can deliver a smaller, tightly defined population | Eligibility is narrow, so criterion-level feasibility matters most |
| Phase III | How many sites, in which countries, over what period | The full funnel with ranges, plus competing trial load by geography |
| Phase IV and post-approval | Which sites can sustain long follow-up | Retention history rather than enrollment speed |
| Site feasibility | Which candidate sites to take forward | Measured funnel stages, not a questionnaire score |
| Investigator selection | Which investigators to approach | Indication experience and research record, not only prior use by you |
| Patient identification | Who might be eligible for pre-screening | Consent scope and permitted use, which govern before any analysis |
| Recruitment planning | Where to place effort and spend | Channels measured on randomisations, not referrals or impressions |
| Enrollment optimization | What to change when the funnel is underperforming | Which stage is losing patients, because the fix differs by stage |
| Representativeness | Whether the enrolled population reflects the treated one | What was actually collected, and the geography of site selection |
| Decentralized and hybrid trials | Whether remote participation widens the reachable population | Which filters it removes and which new ones it introduces |
| Patient engagement and retention | Where retention effort will pay | Early behavioural signals and visit burden rather than late dropout counts |
| Site performance improvement | Which sites to support, and how | Conversion and retention, adjusted, rather than raw enrollment rank |
Before any external data purchase, the highest-value asset for this work is your own history: funnel ratios, screen failure patterns, site conversion and retention across every study you have run. It is almost always scattered across study-specific systems and spreadsheets, never assembled, and discarded at study close.
| Source | What it contributes | The constraint it brings |
|---|---|---|
| Your completed studies | Measured funnel ratios, screen failure, conversion, retention | Study-specific definitions, so standardisation comes before analysis |
| CTMS | Site attributes, activation history, operational performance | Describes sites you used, which is the survivorship problem |
| EDC | Screening, eligibility, randomisation and discontinuation detail | Criterion-level screen failure only if the build captured it |
| IRT | Randomisation and stratification | Treatment-revealing, so funnel views need blinding-aware aggregation |
| eCOA and ePRO | Compliance and engagement behaviour | Compliance decay is a retention signal before it is a data issue |
| Recruitment platforms | Referral, campaign and pre-screening activity | Referral volume is not enrollment, so measure to randomisation |
| Site databases and registries | Investigator experience, prior trials, competing studies | Coverage and currency vary, and absence is not evidence |
| EHR and real-world data | Population size and structured clinical characteristics | Permitted use and consent govern, and collected for care not research |
Keep the funnel data when the study closes. At database lock, the operational and recruitment data for a study stops being maintained and in many organizations effectively disappears, because the next feasibility exercise has no route to it. The fix is a deliberate close-out step: extract the funnel ratios, criterion-level screen failure, site-level conversion, discontinuation timing and the final versus estimated enrollment for every site, and write them into a history layer that is not owned by any one study. It takes a day per study. Over a therapeutic area it is the difference between estimating feasibility and computing it.
Most analytics on this page uses study history and site attributes and raises ordinary governance questions. Patient identification is different in kind, because it means processing identifiable health information about people who have not consented to anything and may never be approached. The lawful routes exist and they are narrow. Your privacy, legal, regulatory and clinical functions determine which applies. We build only inside what they establish.
Frameworks where they apply
- ICH E6 Good Clinical Practice principles for the conduct of clinical research
- GxP where a system supports a regulated activity
- Ethics committee and institutional approvals governing any patient-facing activity
- Applicability and classification determined by your regulatory and quality functions
Privacy
- HIPAA where protected health information is processed, including the basis relied upon
- GDPR and equivalent regimes, including lawful basis and data subject rights
- Consent scope honoured in every downstream use, including secondary analysis
- Data residency and cross-border transfer requirements for each jurisdiction
Minimisation
- De-identification and pseudonymisation applied to the analytical dataset, not only the source
- Aggregate and privacy-preserving approaches where a count answers the question
- Re-identification risk assessed before linking sources that were separately de-identified
- Small-cell suppression in any population or representativeness reporting
Access and integrity
- Role-based access resolved per user at query time, with export treated as part of it
- Study blinding respected in funnel and randomisation-adjacent views
- Lineage from a reported figure back to the source and the extract it came from
- Audit trail over access, definition changes and releases of population figures
Discontinuation is usually reviewed as a cumulative count, months after the decisions that drove it. The behavioural signals that precede it are visible within the first few visits, and that early period is also the only point at which retention effort is cheap and plausible.
| Failure | How it presents |
|---|---|
| Referral volume is reported as recruitment | The top of the funnel is healthy and randomisations are flat |
| One criterion is driving screen failure | Screen failure rate is high and the reason codes are too coarse to show why |
| A site converts poorly but enrols adequately | Volume looks acceptable while it is burning through candidates |
| Sites are ranked without adjusting for activation date | A site live for six weeks is compared with one recruiting for a year |
| Early dropout signals are not watched | Discontinuation appears at month six with no prior indicator |
| Visit burden exceeds what patients will sustain | Retention falls after a specific visit, repeatedly, across sites |
| The site model was never validated | Selected sites underperform and the model is never scored against outcome |
| Representativeness goes unmeasured | No comparison exists because the fields were optional and inconsistently used |
| Competing trials activate mid-study | A site’s rate drops and the cause is outside your study entirely |
The outcome that matters is whether recruitment planning stopped being a negotiation between optimistic numbers and became a calculation with stated sources and ranges. Everything else on this list follows from that.
| Outcome | How to tell |
|---|---|
| Site selection improved | Selected sites hit forecast more often, and the model is scored every study |
| Patient populations identified realistically | Feasibility produces a funnel with ranges and a source per stage |
| Recruitment planning stronger | Channel and site effort allocated on randomisations rather than referrals |
| Avoidable screen failures reduced | Criterion costs were known before the protocol was finalised |
| Retention supported | Early signals reach sites in time to matter, and burden findings reach protocol design |
| Site productivity optimized | Benchmarks measure conversion and retention, and sites accept them |
| Enrollment more predictable | Forecasts carry ranges, are measured against outcome, and error is falling |
We work with pharmaceutical, biotechnology and CRO organizations on site and patient analytics: population funnels modelled multiplicatively with sources and ranges, site scoring that can assess a site you have never used, criterion-level screen failure analysis, early retention signals, and a history layer that outlives the study. Your clinical, medical, operations and privacy functions own the decisions and the patient relationship.
CaliberFocus builds site and patient analytics for pharmaceutical, biotechnology and CRO organizations. We model population funnels multiplicatively with a source and range for every stage, score sites on attributes so a site you have never used can be assessed, analyse screen failure at criterion level, surface early retention signals, and keep a history layer that outlives each study.
- Population funnels with ranges and a stated source per stage
- Site scoring that can assess sites you have never used
- Criterion-level screen failure analysis before protocols are final
- Patient identification only inside the lawful route your teams set
Security & Compliance
