%%{init: {'theme':'base','themeVariables':{'primaryColor':'#28344f','primaryTextColor':'#eef2fb','primaryBorderColor':'#7d9bd6','lineColor':'#aab8d6','clusterBkg':'#161d33','clusterBorder':'#3a476a','tertiaryColor':'#20406a','tertiaryTextColor':'#eef2fb','tertiaryBorderColor':'#7d9bd6','secondaryColor':'#20406a','secondaryTextColor':'#eef2fb'}}}%%
flowchart LR
SRC["Data sources<br/>EDC · safety DB · labs<br/>IXRS · CTMS · finance"] --> RT{"Route by<br/>sensitivity"}
RT -->|"safety / PK /<br/>unblinded"| LOC["On-device model<br/>zero egress"]
RT -->|"non-sensitive<br/>ops"| CLD["Cloud Claude<br/>API + BAA"]
LOC --> ADV["Advanced engine<br/>OT + Active Inference"]
CLD --> ADV
ADV --> SC["Severity-weighted<br/>roll-up to tier RAG"]
SC --> ESC["Alert +<br/>escalation"]
AI for Biometrics
Trial-Termination Early-Warning
A visual, runnable guide to the 3-tier risk console
Run / share / read / build - pick your mode. This page shows the console (diagrams + a live roll-up), the Run & build the Shiny app lets you run it and set a what-if, the dashboard is a shareable snapshot, and the run-and-build page shows how it is made. Synthetic, no PHI, no participant records.
The rule. The AI flags, aggregates, and drafts the alert. The SRC / DSMB / medical monitor make every DLT and safety call; the PM owns the study; the account lead owns the client signal. Reported safety numbers come from the validated safety database, never from a model.
Three tiers, one roll-up
The scoring is transparent - each signal has a level (green/amber/red) and a weight; a tier is the weighted blend, and the overall index weights the tiers (participant safety heaviest).
The live console - computed from the signal feed
Everything below is computed from one file, the signal feed: signals.csv (in the shiny/data/ folder). Edit that CSV (or run the app) and the roll-up, the tiers, and the feed all move.
35 · AMBER
Overall termination risk
38 · AMBER
Participant tier
24 · GREEN
Study tier
44 · AMBER
Client tier
Client · AR-60 Accounts-receivable aging > 60 days
payment-aging KRI breached
Participant · S-1042 Candidate Hy’s-Law / DILI
ALT 3.4x ULN + total bili 2.1x ULN (DLT window, day 9)
Client · COMMS Sponsor comms cadence ADV
-40% month-on-month - behavior drift, relationship watch
Client · SCOPE-CO Scope-reduction change-orders
2 this quarter
Participant · S-1042-ai Free-energy surprise rising ADV
active-inference EFE recommends an unscheduled LFT since day 7
Participant · S-1039 QTcF prolongation +48 ms
watch >30 ms, alert >60 ms
Participant · COH4-AE Within-participant AE cluster
cohort 4 trending Grade 2 to Grade 3
Study · SF-RATE Screen-failure rate
38% vs 25% plan - enrollment KRI breached
Study · LFT-DRIFT Cohort-4 LFT distribution drift ADV
Wasserstein tail mass moving up vs the pooled barycenter
Study · ENROLL Enrollment rate vs plan
-22% vs plan
Study · QRY-AGE Open-query aging
backlog rising ahead of database lock
| Tier | ID | Signal | Category | Engine | Level | Weight |
|---|---|---|---|---|---|---|
| Participant | S-1042 | Candidate Hy’s-Law / DILI | Safety | On-device | red | 10 |
| Participant | S-1042-ai | Free-energy surprise rising | Safety | On-device | amber | 6 |
| Participant | S-1039 | QTcF prolongation +48 ms | Safety | On-device | amber | 7 |
| Participant | COH4-AE | Within-participant AE cluster | Safety | On-device | amber | 6 |
| Participant | PK-EXP | PK exposure vs predicted Cmax | PK | On-device | green | 5 |
| Participant | DLT-TALLY | Cohort DLT tally | Safety | On-device | green | 8 |
| Participant | SAE-NEW | New unexpected SAE | Safety | On-device | green | 9 |
| Study | SF-RATE | Screen-failure rate | Enrollment | Cloud | amber | 7 |
| Study | LFT-DRIFT | Cohort-4 LFT distribution drift | Safety | On-device | amber | 6 |
| Study | ESC-33 | Cohort-4 escalation (3+3) | Dose-escalation | Cloud | green | 8 |
| Study | QTL-SAE | Safety QTL (SAE rate) | Safety | Cloud | green | 8 |
| Study | ENROLL | Enrollment rate vs plan | Enrollment | Cloud | amber | 5 |
| Study | QRY-AGE | Open-query aging | Data | Cloud | amber | 4 |
| Study | PROTDEV | Protocol-deviation rate | Compliance | Cloud | green | 5 |
| Study | SITE-ACT | Site-activation lag | Operations | Cloud | green | 3 |
| Client | AR-60 | Accounts-receivable aging > 60 days | Finance | Cloud | red | 9 |
| Client | COMMS | Sponsor comms cadence | Relationship | Cloud | amber | 5 |
| Client | SCOPE-CO | Scope-reduction change-orders | Commercial | Cloud | amber | 6 |
| Client | RENEWAL | Contract-renewal signal | Commercial | Cloud | green | 5 |
| Client | FILINGS | Public-filings sentiment | Commercial | Cloud | green | 4 |
| Client | SATIS | Sponsor-satisfaction pulse | Relationship | Cloud | green | 4 |
Run it - and change it
shiny::runApp("risk_monitor/shiny")
# needs: install.packages(c("shiny","bslib","DT"))Select a signal, set its level, and watch the tier and overall risk re-colour. The run-and-build page shows how the console is built from R/logic.R.
Going live: the feed file signals.csv is a representative subset of the 65-signal catalog. Replace it with your KRI / stopping-rule feed (same columns) and the roll-up is yours. Participant-tier signals are de-identified states, never records.
The other formats
| Mode | Artifact |
|---|---|
| Run | Run & build the Shiny app |
| Share (no R) | Dashboard |
| Build | Run-and-build walkthrough |
Synthetic, no PHI. The AI flags; the SRC / DSMB / PM / account lead decide.