# Measuring AI impact after deployment
A hospital in Rotterdam deployed an AI tool to flag suspected strokes on CT scans. In month one, radiologists loved it. By month eight, the flag rate had quietly crept up by 30 percent, and nobody knew if the tool got better or if it started crying wolf. That gap between "we launched it" and "we know it works" is where most biotech and medtech AI value leaks out.
This lesson gives you the metric discipline to close that gap: what to measure, when signals appear, and how to catch performance drift before it hurts patients or budgets.
The core distinction: leading metrics move early and predict impact. Lagging metrics confirm value but arrive late.
You need both. Leading metrics let you course correct. Lagging metrics let you prove ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → (return on investment) to the people who signed the check.
Drug discovery (screening).
A team using AI-driven virtual screening might cut a screening cycle from 12 weeks to 6. That is a leading signal. Whether those hits survive lab validation, that is the lagging truth. AI can inflate the leading number by proposing many plausible but weak candidates.
Radiology.
Clinical trials.
AI patient-matching tools promise faster recruitment. The leading metric (enrollment rate) can look great while the lagging metric (screen-failure rate) reveals the model is matching the wrong people.
You cannot measure impact without a counterfactual: what would have happened without the AI.
Three practical options, weakest to strongest:
1. Historical baseline. Compare to the six months before deployment. Cheap, but confounded by seasonality and workflow changes.
2. Parallel cohort. Some sites or readers use AI, others do not, during the same period.
3. Randomized comparison. Cases randomly routed with or without AI. Strongest, rarely feasible in live clinical settings.
For a radiology AI, a common design: run silent mode first. The model predicts, but clinicians never see the output. You collect performance data against real outcomes with zero clinical risk. Only then flip to live use. Silent mode is your honest baseline.
Keep this strictly about AI-attributable value, not general accounting.
Suppose a radiology group processes 40,000 chest CT studies per year. An AI triage tool credibly saves an average of 3 minutes of radiologist time per study through better prioritization and pre-population of findings.
Now subtract AI costs: per-study licensing, integration, and monitoring. If licensing is 4 US dollars per study, that is 160,000 US dollars per year, plus say 40,000 for monitoring and oversight.
Caveat: time saved only converts to value if it is redeployed (more reads, less overtime). If radiologists simply have marginally lighter days, the cash impact is near zero. Always state your realization assumption. This is where inflated AI business cases collapse.
A model that passed validation in 2025 is not guaranteed to perform in 2026. This is drift, and it comes in flavors:
Drift is not hypothetical. When a new imaging device rolls out, an AI trained on the old one can silently degrade. Your monitoring must catch this before a clinician does.
Track three layers continuously:
1. Input distribution. Are incoming images, lab values, or patient features statistically similar to the training set?
2. Prediction distribution. Is the model's positive rate stable? A jump from 8 percent to 15 percent flagged cases needs explanation.
3. Outcome performance. Where ground truth is available (biopsy result, confirmed diagnosis, trial enrollment success), track accuracy over time. This lags, but it is the truth layer.
A lightweight statistical check for input drift, using the Population Stability Index (PSI), a standard measure of how much a distribution has shifted:
import numpy as np
def psi(expected, actual, bins=10):
breakpoints = np.percentile(expected, np.linspace(0, 100, bins + 1))
breakpoints[0], breakpoints[-1] = -np.inf, np.inf
e = np.histogram(expected, breakpoints)[0] / len(expected)
a = np.histogram(actual, breakpoints)[0] / len(actual)
e, a = np.clip(e, 1e-6, None), np.clip(a, 1e-6, None)
return np.sum((a - e) * np.log(a / e))
# Rule of thumb: PSI < 0.1 stable, 0.1-0.25 moderate shift, > 0.25 significant driftRun this weekly on a key input feature. A PSI above 0.25 is your early warning to investigate before outcomes degrade.
🎬 [VIDEO: "Monitoring Machine Learning Models in Production" - youtube.com - practical walkthrough of drift detection and monitoring pipelines for deployed models]
If your AI is a medical device, monitoring is not optional. In the US, the FDA (Food and Drug Administration) regulates AI-based SaMD (Software as a Medical Device). In 2025 the FDA finalized guidance on the Predetermined Change Control Plan (PCCP), which lets you pre-specify how a model may be updated and monitored without a new submission each time. Read the FDA's overview of AI-enabled medical devices.
In Europe, AI medical devices fall under the MDR (Medical Device Regulation) and, increasingly, the EU AI Act, which classifies most medical AI as high risk and mandates post-market monitoring and human oversight. Your drift dashboards are not just good practice; they are becoming evidence for regulators.
Key point: post-market surveillance and your performance monitoring are the same activity. Design them together.
Vérification des acquis
1. What is the fundamental distinction between leading and lagging metrics when measuring AI impact after deployment?
2. In the Rotterdam stroke-detection example, the flag rate rose 30 percent over eight months without anyone knowing whether the tool improved or started 'crying wolf.' What core concept does this scenario illustrate?
3. A drug-discovery team reports that AI virtual screening cut a screening cycle from 12 weeks to 6. Why should this improvement be interpreted cautiously?
4. Select ALL correct answers. Which of the following are correctly categorized as LAGGING metrics in the examples given?
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers. Why does an organization need BOTH leading and lagging metrics rather than just one type?
Sélectionnez toutes les réponses correctes.
Vanity leading metrics. "The model achieves 94 percent AUC" (AUC, or Area Under the Curve, measures how well a model separates positive from negative cases). Impressive in the lab, meaningless if throughput and outcomes do not move in the clinic.
Automation bias. Once clinicians trust the AI, they may stop double-checking. Your accuracy metrics can look stable while human oversight quietly erodes. Track override rates and how often clinicians agree with the AI. A sudden drop in overrides can signal over-reliance, not better performance.
Attribution error. Recruitment sped up, but was it the AI or a new coordinator? Without a counterfactual, you are guessing. Guessing gets your budget cut at the next review.
Survivorship in discovery. In screening, celebrate validated hits, not proposed ones. AI that floods the pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète → with weak candidates looks productive and wastes lab time.
Assign an owner. A dashboard nobody reads is drift waiting to happen.