Gains:
- Ability to interpret the concepts of sensitivity, specificity, false negative and false positive in the context of radiology and to critically read the metrics of a model.
- Being able to recognize automation bias (over-reliance on artificial intelligence) and, on the contrary, alarm fatigue, and protect the reading discipline against these two traps
- Understanding that the radiologist must read an area that is not marked by artificial intelligence, and that a negative AI output is not a guarantee of diagnosis.
The two most important numbers for a radiology AI are how many findings it catches and how many false alarms it raises. If a manufacturer tells you “our model is 96% accurate,” that alone says almost nothing. Because for a rare finding (say, 10 diseases in 1,000 exams), even a model that flags nothing may appear to be “99% accurate” — it correctly recognizes 990 healthy ones and misses only 10 patients. In this unit, we'll learn how to critically read radiology's crucial metrics—sensitivity, specificity, false negative, false positive—like an expert, and recognize two dangerous human traps—automation bias and alarm fatigue.
Core principle: An area not marked by artificial intelligence is also read by the radiologist. A negative AI result is not a guarantee of diagnosis; the model may have missed the finding (false negative). The raison d'être of human monitoring is to capture the exact moments when the model is safely wrong.
Reading metrics like an expert
Let's clarify four basic concepts. Let's say we tested a model on 1000 exams and 100 of them actually had the disease.
- True positive (TP): The model marked the patient, truly sick.
- False negative (FN): The model did not mark but the patient is ill — miss. The most dangerous in radiology.
- False positive (FP): The model flagged but the patient is healthy—false alarm.
- True negative (TN): Model did not mark, really healthy.
Two basic metrics arise from these:
- Sensitivity = TP / (TP + FN): How many real patients did we catch? High sensitivity means low abduction.
- Specificity = TN / (TN + FP): How many of the healthy did we count as healthy? High specificity means fewer false alarms.
metric
formula
When it's high
Radiological significance
Sensitivity
TP/(TP+FN)
Few false negatives
Critical to avoid missing (screening, triage)
specificity
TN/(TN+FP)
Few false positives
Reduces unnecessary examination
False negative rate
FN/(TP+FN)
bad
Silent harm; radiologist must catch
False positive load
number of FPs
load increases
Alarm fatigue, recall
"accuracy"
(TP+TN)/total
misleading
Appears high in rare disease, deceives
A model has a threshold (above what the score is considered “positive”), and this threshold changes sensitivity and specificity in opposite directions: if you lower the threshold, you catch more (sensitivity ↑) but raise more false alarms (specificity ↓). This is not an engineering setting, but a clinical decision: the heavy cost of missing, or the burden of a false alarm? Sensitivity is generally prioritized in screening and triage.
Caution: Never trust a single number like "95% accuracy". In rare findings, accuracy is misleading. The true performance of a model cannot be known without asking the sensitivity, specificity, false negative rate, and the population in which it was measured.
Two human traps
Automation bias: When people over-rely on the output of an automated system and relax their own independent judgment. If the radiologist superficially skips the image by saying "AI said it was negative, so it is clean", the finding missed by the model will be completely skipped. When false negatives combine with automation bias, the raison d'être of human inspection is eliminated.
Alert fatigue: As a result of the model producing a large number of false positives, the radiologist loses confidence in the flags and begins to reflexively eliminate them all. A true finding after the tenth false alarm may also be eliminated. These two traps are in opposite directions, but their results are the same: missing a real finding.
The antidote to these two traps is the same: no matter what the AI output says, the radiologist does his own systematic reading. Whether the AI marks it or not, the entire image is scanned. AI is a “second eye”; The first eye is always the radiologist.
three mini cases
Case 1 — False negative + automation bias. A chest radiograph CAD misses (false negative) a small nodule in the upper left zone. On a busy day, the radiologist rushes through the radiograph thinking “CAD is clear” (automation bias). The nodule is noticed after 8 months as a mass of 2 cm in size. Two errors combined: model missed, human did not check. Lesson: despite negative CAD, systematic reading is essential.
Case 2 — Alarm fatigue. A pneumothorax model throws an average of 8 flags per day for two weeks, only 1-2 of which are real (about 80% false positives). The radiologist loses confidence in the flags. In the third week, a true apical pneumothorax flag is also reflexively passed as "no"; The patient gets worse after a day. Lesson: models with a high false positive rate should be monitored, threshold/model reviewed, and each flag confirmed anyway.
Case 3 — The right balance. In a stroke center, the brain CT triage model works with high sensitivity; False positives are high, but the radiologist confirms each flag in 60 seconds and detects real bleeding within minutes. The team monitors the false positive rate weekly, reviewing the threshold with the manufacturer when it exceeds 70%. Here, metrics were managed consciously; Neither automation bias nor alarm fatigue has set in.
Weak prompt / Strong prompt
Weak prompt:
This model is 97% accurate, is it reliable?
Single metric is misleading; cannot be answered without asking questions about population, sensitivity, specificity, and false negatives.
Powerful prompt:
Your role: HELP me to critically read the performance metrics of an AI model.Decision making; Explain what questions I should ask and what the answers mean. I will give you the claims of a model. Consider: (1) which metrics are given and which are missing (sensitivity, specificity, false negative rate, population, device), (2) whether “accuracy” alone is misleading, (3) whether it is appropriate to use this model for triage/screening or diagnosis. The final decision is up to the radiologist/institution. Claim: "Model tested in 1200 patients with 97% accuracy."
Strong demand calls into question missing metrics and breaks the reliance on single numbers.
Copiable prompt templates
METRIC CRITICAL READING TEMPLATEYour role: HELP. I'll give you a model's performance claim. List missing metrics (sensitivity, specificity, false negative rate, test population, device/protocol) and where a single number (accuracy) may be misleading. Discuss for which task (triage/screening/diagnostic-assistance) the model is appropriate to use. The decision is in the institution. Claim: [write]
THRESHOLD BALANCE DESCRIPTION TEMPLATEI will give you a scenario of raising/lowering the threshold of a model. Describe the impact of each scenario on sensitivity, specificity, false negative, and false positive and write down its clinical outcome (miss vs. unnecessary recall). The decision is clinical/institutional. Script: [write]
AUTOMATION BIAS SELF-AUDIT TEMPLATEProduce me a checklist that will protect my reading discipline against automation bias: what steps should I take even with negative AI output, how to maintain systematic scanning, how to keep AI as an "assistant" rather than a "delivery tool".
ALERT FATIGUE MONITORING TEMPLATEI will give you weekly flag counts and actual positive numbers. Calculate the false positive rate, assess the risk of alert fatigue, and suggest at what threshold I should take action to revise the threshold/model. Remind me of the principle that every flag should still be confirmed. Data: [write]
Common mistakes
- Relying on the single "truth" number. The rare finding is also misleading; Sensitivity and specificity should be asked together.
- Treating negative AI output as a diagnostic guarantee. The model may have missed it; Systematic reading is a must.
- Falling into automation bias. Over-reliance on AI undermines independent judgment.
- Ignoring alarm fatigue. High false positives should be monitored; each flag must still be confirmed.
- Mistaking the threshold for a "technical setting". The threshold is a clinical balance; The radiologist/institution weighs the cost of misses and false alarms.
Tip: Make a rule for yourself: "No matter what the AI says, I read the entire image." This single rule simultaneously curbs both automation bias (not relaxing on the negative) and alarm fatigue (not reflexively eliminating the positive).
In summary
Evaluating a radiology model is not about looking at a single “accuracy” number. Sensitivity (no misses), specificity (no false alarms), false negative rate, and test population should be read together; The choice of threshold is a clinical balance. The two human traps — overconfidence in AI (automation bias) and getting used to false alarms and eliminating the flag (alarm fatigue) — go in opposite directions but end with the same result, missing a real finding. There is only one antidote: No matter what the AI says, the radiologist makes his own systematic reading. Negative AI is not a guarantee, but at most a helpful signal; The unmarked area must also be read.
Application task
Take the promotional text of a model you have (or a sample). With the “Metrics Critical Reading” template, evaluate which metrics are provided, which are missing, and for what task it is appropriate to use the model. Then design a simple chart to track how many times a triage flag comes true over the course of a week; Write down what action you will take if the false positive rate exceeds the threshold you set (for example, 70%). Finally, adapt the “Automation Bias Self-Audit” list into your own reading routine.
checklist
- [ ] I didn't rely on the single "accuracy" number; I asked about sensitivity and specificity.
- [ ] I learned the false negative rate and test population.
- [ ] Despite the negative AI output, I did my systematic reading.
- [ ] I wasn't overly confident in AI (versus automation bias).
- [ ] I watched for false positives; I confirmed every flag (against alert fatigue).
- [ ] I took into account that the choice of threshold is a clinical balance.
- [ ] I followed the rule of "I read the entire image no matter what the AI says."