How forecasted patient demand becomes a supply response
A clinical-supply plan fails quietly. The enrollment curve moves, the kit requirement moves with it, and by the time a depot notices its coverage has thinned, the only remaining options are expensive ones — an air-freight lane, a borrowed batch, or a site that pauses dosing.
This experience owns the segment of the family where a modelled enrollment outlook stops being a forecast and starts being a supply obligation. It translates patient demand into kit demand, reconciles that demand against depot and site inventory, surfaces the stockout and expiry exposures competing for the same kits, and puts forward a resupply response with its reasoning attached — for a person to approve.
Scope Within the Family
This experience begins where the enrollment outlook ends. The forecasted patient trajectory arrives as an input to be translated into supply requirement — not as something this page models, evaluates or defends.
This experience owns
This experience does not own
Those capabilities belong to Clinical Enrollment Forecasting & Site Risk Intelligence and Clinical Trial Site Performance Intelligence.
A note on honesty
The prototype behind this page does contain an internal enrollment projection, because it needs one to derive kit demand. That projection is used here purely as a supply input. Its methodology, its evaluation and any judgement about how far to trust it are deliberately not presented on this page — they are the subject of the sibling experience, and duplicating them here would blur a boundary the family depends on.
The Premise
The hard part is not calculating a quantity. It is that the quantity depends on a derived demand signal, a distributed stock position and an expiry clock — three things that are usually measured separately and reconciled late.
No system measures how many kits a trial will need. It is derived from an enrollment outlook, a randomization ratio and a visit schedule — so the demand number is only ever as good as the forecast behind it.
The batch that best covers a shortage may also be the batch closest to expiry. Coverage and waste are the same decision viewed from two sides, and neither is safe to optimise alone.
A site runs thin while a depot two countries away sits long. Seeing both positions in one place is what turns a stockout alert into a resupply option.
Nothing here issues a shipment, posts a transaction or releases a batch. Every output is a proposal that stops at a qualified planner.
Business Context
The failure is rarely a missing report. It is a demand signal that arrived late, a shortage ranked without a quantity, and a proposal whose reasoning did not travel with it.
The planning problem
Each signal is available somewhere. What is missing is a place where they are held against each other early enough for an ordinary intervention to still be an option.
The connected supply response
This is not a supply planning system, an optimiser or an execution platform. It demonstrates a governed path from a forecasted demand signal to a proposal a person still has to accept.
Reading this page correctly
Everything shown is generated synthetic data. Trial, site, country, depot, batch, kit, carrier and patient identifiers are generated synthetic labels — the patient identifier field in the underlying IRT/RTSM dataset is explicitly named as a fake, and values such as “Emergency Unblinding Placeholder” are placeholder transaction-type strings in a synthetic categorical column, not a capability this prototype offers or claims. Monetary figures are illustrative synthetic exposures rather than financial estimates, and no forecast-accuracy or model-quality figure is published anywhere on this page.
Decision Framework
Nine stages, in order. Position is deliberately established before response, and expiry exposure is carried alongside shortage rather than after it — that ordering is what keeps a resupply proposal reviewable.
Functional Coverage
The prototype contains fifteen screens running against one shared synthetic dataset. Eight are presented as motion showcases below; all fifteen are listed here so the functional breadth is visible rather than implied.
Showcase 01 · Portfolio position
Where this sits in the journey
The opening position across the synthetic portfolio, before any single exposure is drilled into. Every figure here is an aggregate of the same dataset the later screens score in detail.
Business question
Across every active trial, where is clinical supply under the most pressure right now — and is that pressure a demand problem, an inventory problem or an expiry problem?
Recorded from the running prototype on synthetic data. Trial, country and depot identifiers are generated labels, not real studies, markets or facilities.
Key signals on the screen
Decision supported
Which trials, countries and depots deserve a planner's first hour — and which exposure type is driving the pressure before any drill-down is opened.
Captured synthetic state
12 active trials, 233 active sites, 8,611 critical stockout rows, 5,083 rows below buffer, 1,869 depot shortage alerts, USD 384,141,948 batch expiry value at risk, 2,263 shipments at delay risk, 768 cold-chain risk shipments, patient dosing continuity risk 35/100, 81 open critical exceptions and 10,119 recommended SAP actions.
Interpretation & review boundary
A triage view, not a status of record. The counters describe a synthetic dataset at a point in time and carry no operational, clinical or regulatory standing.
Showcase 02 · Consumed input becomes demand
Where this sits in the journey
This is the handover point in the family. The enrollment trajectory enters as an input; what this experience owns is the translation of that trajectory into a kit requirement.
Business question
Given the enrollment outlook we have been handed, how many kits of each type does this trial actually need over the forecast horizon?
Recorded from the running prototype on synthetic data. The enrollment curve shown here is a consumed planning input, not a forecasting capability this experience owns.
Key signals on the screen
Decision supported
How much of each kit type to plan for, and which kit types carry enough demand concentration to be worth managing separately.
Captured synthetic state
A forecast active-patient band of roughly 22,000–24,000 across a twelve-month horizon spanning 2026 and 2027, resolving into per-kit-type demand of roughly 4,000–12,000 units across ten kit types including Auto-Injector, Cold-Chain-Vial, Pre-Filled-Syringe and four oral solid presentations.
Interpretation & review boundary
Forecast methodology, rolling-origin evaluation and forecast-model comparison are deliberately absent here — they belong to the Enrollment Forecasting experience. No forecast-accuracy claim is made on this page.
Showcase 03 · Where the exposure lands
Where this sits in the journey
Demand has become a requirement; this screen establishes where that requirement is not currently covered. Risk is banded rather than reported as a single ranked list.
Business question
Which sites are close enough to running out that dosing continuity is threatened — and how large is the gap at each of them?
Recorded from the running prototype on synthetic data. Site and country identifiers are generated labels; no real investigator site or market is represented.
Key signals on the screen
Decision supported
Which sites need intervention first, and whether the right response is a site-level resupply or a country-level rebalancing.
Captured synthetic state
25,000 scored rows with 8,611 in the critical tier and 137 in the high tier, an average of 231.0 days of supply remaining, and a top-shortage list peaking at roughly 8,000 kits per site. The chart's own axis labelling indicates counts are held at site-and-kit granularity rather than distinct-site granularity.
Interpretation & review boundary
Site operational performance, dropout risk, protocol-deviation risk and site selection are not analysed here — those belong to Site Performance Intelligence. This screen scores supply exposure only.
Showcase 04 · The supply position
Where this sits in the journey
Before a resupply proposal can be read as sensible, the network position has to be established: which depots can answer a shortage, and which are short themselves.
Business question
Where is stock actually sitting relative to forecast country demand, and which depots are too thin to be a source rather than a destination?
Recorded from the running prototype on synthetic data. Depot identifiers are generated labels and do not correspond to any real distribution facility.
Key signals on the screen
Decision supported
Which depots are candidates for inbound transfer, and which can safely act as a source for a downstream site resupply.
Captured synthetic state
12 depots scored, days of supply ranging from roughly 6.6 to 7.4 with DEP-003 lowest and DEP-007 highest, composite depot health scores clustered around 20–21.5, and per-depot demand-versus-stock comparison in the low millions of kits.
Interpretation & review boundary
A coverage and balance view on synthetic stock. It is not an inventory system of record, performs no allocation, and reserves, blocks or releases nothing.
Showcase 05 · The competing constraint
Where this sits in the journey
Expiry exposure is carried in parallel with shortage exposure rather than after it, because the same batch can be the answer to one and the cause of the other.
Business question
Which batches are likely to expire before they are consumed, and how much value is sitting behind that risk?
Recorded from the running prototype on synthetic data. Batch identifiers and monetary values are generated for demonstration and are not financial estimates.
Key signals on the screen
Decision supported
Which batches to consume, redeploy or plan around first — and whether a shortage should be answered from stock that is itself close to expiring.
Captured synthetic state
1,000 batches scored, USD 384,141,948 of value at risk, and 91 batches in the critical band, with the risk-score distribution concentrated below the mid-range and a visible upper tail.
Interpretation & review boundary
The monetary figure is an illustrative synthetic exposure, not a forecast, a valuation or a financial estimate. Nothing here performs GMP disposition, quality release or quarantine.
Showcase 06 · The modelled response
Where this sits in the journey
This is the response the whole page has been building toward — and it is deterministic rules applied to model-scored risk, not a mathematical optimiser. The method note below the showcases states exactly how it works.
Business question
Given the exposures established above, what resupply is proposed, how urgent is each proposal, and how should it move?
Recorded from the running prototype on synthetic data. The screen is named Resupply Optimization inside the application; the underlying logic is threshold and rule based, as described in the method note below.
Key signals on the screen
Decision supported
Which resupply proposals a planner should review first, at what quantity, and with what transport urgency — before anything is approved.
Captured synthetic state
8,781 recommendations, of which 8,634 sit in the critical urgency band and 113 in high, aggregating to roughly 40,000–60,000 recommended kits per country across the top fifteen countries.
Interpretation & review boundary
Recommendation only. The application states on this screen that no SAP transaction is executed and no autonomous action is taken, and that planner review and approval are required before any resupply is issued.
Showcase 07 · Replanning
Where this sits in the journey
Three things stay separate: the retained baseline, the declared scenario assumption, and the projected result that follows from it. The projected result is never presented as the accepted position.
Business question
If enrollment or supply conditions change, how far does the supply position move — and is the change large enough to act on now?
Recorded from the running prototype on synthetic data. Scenario outcomes are modelled illustrative what-ifs, not predictions of a real trial or supply network.
Key signals on the screen
Decision supported
Whether a change in enrollment or supply conditions justifies pre-emptive resupply now, or whether the position absorbs it.
Captured synthetic state
A baseline of 35.1 site stockout risk, 79.3 depot supply risk and USD 384,141,948 expiry value at risk; under an 'Enrollment +30%' assumption, site stockout risk moves to 41.8 (+6.7 points) while depot supply risk and expiry exposure hold flat, with patient dosing continuity risk at 31.9/100.
Interpretation & review boundary
Scenario deltas are applied to baseline positions as declared assumptions. They are not probabilistic forecasts, not confidence-bounded, and carry no operational commitment.
Showcase 08 · The governed handoff
Where this sits in the journey
The end of the chain. Proposals are expressed in SAP-oriented action vocabulary so a reviewer can see what would be actioned where — and the screen is explicit that nothing is.
Business question
If a planner accepted these proposals, what would actually be actioned, in which process area, and of what type?
Recorded from the running prototype on synthetic data. This is an SAP ICSP-inspired vocabulary demonstration and is not affiliated with, endorsed by, or connected to SAP.
Key signals on the screen
Decision supported
Which function owns each proposed action, and what the review queue looks like before a qualified planner accepts or rejects any of it.
Captured synthetic state
10,119 recommended actions distributed across S/4HANA, TM, QM, ICSP and EWM process areas, with S/4HANA carrying the large majority, split across Exception Handling, Resupply and FEFO action types.
Interpretation & review boundary
SAP-oriented naming is a vocabulary alignment exercise on synthetic data. There is no SAP connection, no transaction posting, no interface and no execution of any kind.
Method, Stated Plainly
The screen inside the application is named Resupply Optimization. The logic behind it is deterministic rule-based selection applied on top of model-scored risk. It is described that way here because that is what it is.
Where the prototype’s own in-application glossary describes resupply in cost-minimising terms, that wording is not reproduced here — the implemented logic contains no cost objective, and the source behaviour is the authority for what this page claims.
Supporting Intelligence
These are not presented as motion showcases, but each is fully implemented and runs against the same synthetic dataset as the eight views above.
A locally simulated, rule-based and retrieval-style assistant that answers supply questions by explaining outputs the models already produced. It is a conversational surface over existing results, not a live language-model integration and not a decision-maker.
Shipments scored for delay probability and cold-chain exposure, with carrier-level risk ranking and recorded temperature-excursion counts by risk level. Presented here as supporting context; deep cold-chain quality intelligence sits elsewhere in the lab.
Open exceptions scored into a composite priority and banded across critical, high, medium and low, so the queue in front of a planner is ordered by impact. The captured state holds 81 critical, 283 high, 465 medium and 171 low exceptions.
An in-application catalogue of the extension modules and architecture layers that sit beyond a baseline planning view, kept visible so the scope of what was actually built is inspectable rather than asserted.
Two explorers that expose the underlying synthetic datasets directly — clinical-supply objects on one side, SAP-style objects on the other — so any figure shown elsewhere in the prototype can be traced back to the rows that produced it.
A plain-language glossary of the concepts the prototype uses, from enrollment-driven kit demand and stockout probability through depot inventory health, FEFO and dosing continuity risk.
Capabilities
Model-based layers and deterministic layers are named separately, because the difference changes how far an output should be trusted.
What it does
Converts a consumed enrollment outlook into per-kit-type demand across trials, arms and visit schedules.
Decision supported
How much of each kit type to plan for over the forecast horizon.
Evidence used
A gradient-boosting regression model over synthetic enrollment, randomization and visit-schedule features.
What it does
Scores and bands sites by stockout risk against days-of-supply remaining and attaches a shortage quantity.
Decision supported
Which sites need intervention before dosing continuity is threatened.
Evidence used
A random-forest classifier over synthetic site inventory, consumption and coverage features.
What it does
Compares depot stock against forecast country demand and scores each depot with a composite health measure.
Decision supported
Which depots can source a resupply and which need inbound transfer themselves.
Evidence used
A deterministic composite score that penalises both shortage and overstock, clipped to a 0–100 range.
What it does
Scores batch-level expiry risk and aggregates the value sitting behind it.
Decision supported
Which batches to consume or plan around before they become waste.
Evidence used
A gradient-boosting regression model over synthetic batch age, shelf-life and consumption features.
What it does
Selects qualifying sites above a risk threshold, bands them by urgency and assigns a quantity, source depot and transport mode.
Decision supported
What resupply to put in front of a planner, and in what order.
Evidence used
Deterministic threshold and rule logic applied on top of model-scored stockout risk — no solver, no cost objective.
What it does
Scores shipments for delay probability and cold-chain exposure and ranks carriers by combined risk.
Decision supported
Which lanes and carriers warrant reroute or closer monitoring.
Evidence used
A random-forest classifier over synthetic shipment, lane and temperature-excursion features.
What it does
Combines impact signals into a single exception priority score and bands the resulting queue.
Decision supported
Which open exception a planner should work next.
Evidence used
A composite priority score with a gradient-boosting model fitted over the same synthetic exception features.
What it does
Applies a declared assumption to retained baseline positions and reports the resulting deltas with a written interpretation.
Decision supported
Whether a change in conditions justifies pre-emptive action now.
Evidence used
Deterministic deltas applied to baseline risk and exposure measures, with a named planner response per scenario.
What it does
Expresses accepted-for-review proposals in SAP-oriented action vocabulary grouped by process area and action type.
Decision supported
Which function owns each proposal, before any human approval.
Evidence used
A deterministic mapping from scored exposures to SAP-ICSP-style action names on synthetic data.
Model families are named because they are visible in the prototype’s own model registry. No accuracy, error or performance figure is published, and none should be inferred from the presence of a model.
System Design
A seeded pipeline from synthetic generation through scoring and deterministic response logic to a review surface, with the model layer and the rule layer kept structurally separate.
Built in Python with Streamlit, pandas, NumPy, scikit-learn and Plotly, running locally against generated synthetic datasets. The SAP ICSP vocabulary is an alignment exercise only: there is no SAP connection, no connector, no interface and no transaction of any kind, and the prototype is not affiliated with or endorsed by SAP.
Lab Standards
Clinical supply sits close to patient dosing, so the boundaries are stated on the screens themselves as well as here. In a domain where an automated action could interrupt a patient’s treatment, the advisory-only policy is the design rather than a limitation.
Honest Status
The prototype runs locally against seeded synthetic data. It is not deployed, not published as an application, and not available to run.
These are deliberate boundaries of a synthetic-data prototype, not pending features. The absence of an optimiser is a statement of fact about the implementation; the advisory-only execution policy is a governance decision.
What This Demonstrates
Building this experience exercised the following, end to end.
Separating a clinical-supply problem into demand translation, coverage position, competing exposures and response — and keeping enrollment forecasting outside the boundary on purpose.
Naming which layer is a supervised model and which is deterministic rule logic, and refusing to describe threshold-based candidate selection as optimisation.
Designing a path that ends at a human reviewer, with the recommendation-only boundary carried on the screens themselves rather than only in a footnote.
Mapping outputs onto SAP-ICSP-style process areas and action types so the work is legible to an enterprise audience without claiming any system connection.
This experience is one of four in the Clinical Trial Supply Chain Intelligence family. The enrollment outlook it consumes is owned by the forecasting sibling, and the site execution behind that outlook is owned by the site experience.