Artificial Intelligence Exhibits Critical Blind Spots in Allergy Test Room Safety Design

We report a pilot study that revealed concerning blind spots in the conceptualization of medical safety requirements by artificial intelligence (AI).

During routine skin testing, a patient experienced anaphylaxis. Although the event was managed effectively, crucial time was lost while staff searched for a pulse oximeter that was stored away from the bedside. This incident highlights global variability in safety practices across allergy testing facilities, despite international guidelines that mandate the immediate availability of emergency equipment [1]. We therefore explored whether AI could contribute to standardizing safety assessments of allergy testing rooms.

We conducted a three-phase pilot study evaluating AI’s baseline understanding of allergy test room requirements and its potential utility as a safety assessment tool following expert training.

• 1) Phase 1: Using GPT-4o’s integrated image generation, we prompted, “Generate a photorealistic image of an allergy test room in a modern outpatient clinic, showing the full room layout.” A single image was generated using GPT-4o (May 2024 version) with default parameters (temperature = 1.0, no seed specification) (Figure 1). This zero-shot approach intentionally avoided iterative refinement in order to capture the AI’s baseline conceptualization. For reproducibility, the same prompt consistently produces images with similar omissions of safety equipment, although specific aesthetic details vary.

• 2) Phase 2: Twelve board-certified allergists (mean experience 5.5 years, range of 1–20) evaluated the AI-generated image using a 22-item checklist derived from international guidelines. Items spanning six domains—emergency equipment, storage, waste management, monitoring, testing infrastructure, and digital systems—were scored on a 0–3 scale (0 = absent, 3 = optimal) [1,2].

• 3) Phase 3: We synthesized expert feedback with comprehensive international guidelines (World Allergy Organization, European Academy of Allergy & Clinical Immunology, American College of Allergy, Asthma & Immunology) and Turkish National Standards (TNASCI) to develop “RoomReady: Allergy Setup Checker,” a publicly available CustomGPT tool [1–4]. RoomReady was trained on more than 500 requirements synthesized from four guidelines, structured through a 22-item expert rubric, and further refined using 20 test room photographs. Validation was performed using four independent clinics not included in training. To assess the quality and accuracy of RoomReady’s outputs, four blinded allergy specialists evaluated the AI-generated assessments using a modified DISCERN instrument (mDISCERN), a validated tool for evaluating the quality of written health information. The 8-item mDISCERN was scored on a 5-point Likert scale (maximum 40 points), with quality categorized as good (32–40), fair (16–31), or poor (8–15) [ 5]. This approach aligns with emerging AI applications in medical facility design [6,7]. For initial proof-of-concept testing, we tested the tool on photographs from two allergy clinics using the prompt, “Please assess this room.”

Figure 1.

Photorealistic allergy test room layout generated by GPT-4o in a zero-shot prompt without human guidance.

Expert evaluation revealed catastrophic safety deficiencies (Table 1). Seven critical items received unanimous zero ratings: oxygen delivery system, vital signs monitor, medical refrigerator, pulse oximeter, clock or timer, allergen extract storage, and skin test application table. This unanimity across all 12 experts was striking. Inter-rater agreement was substantial, suggesting that these omissions were obvious rather than subtle.

Table 1.

Expert assessment of the AI-generated allergy test room (n = 12)

Equipment category and items	Median score	Mode	Range	Experts rating 0 (%)
Emergency equipment
Emergency cart/crash cart	1.5	1	0–3	8.3
Oxygen delivery systema	0	0	0–0	100.0
Vital signs monitora	0	0	0–0	100.0
Examination bed	2.5	3	1–3	0.0
Storage and repository
Medical refrigeratora	0	0	0–0	100.0
Disposable materials storage	0	0	0–1	66.7
Safety and waste management
Sharps container	1.0	0	0–2	33.3
Medical waste container	0	0	0–1	91.7
Hand hygiene station	1.5	1	0–3	8.3
Patient monitoring
Blood pressure monitorb	0	0	0–3	91.7
Pulse oximetera	0	0	0–0	100.0
Clock/timera	0	0	0–0	100.0
Testing infrastructure
Allergen extract storagea	0	0	0–0	100.0
Test materials cabinet	0	0	0–2	66.7
Skin test application tablea	0	0	0–0	100.0
Clean preparation surface	1.5	3	0–3	25.0
Digital systems
Computer	2.0	2	0–3	16.7
General evaluation
Room size & layout	2.0	2	1–2	0.0
Lighting & visibility	2.0	2	1–3	0.0
Air quality & ventilation	2.0	2	1–2	0.0
Patient flow & accessibility	2.0	2	0–2	8.3
General usability for allergy testing	1.0	1	0–2	16.7

AI: artificial intelligence.

Scoring: 0 = completely absent; 1 = present but inadequate; 2 = adequately present; 3 = optimally present.

Composite emergency preparedness score was 33% and critical monitoring infrastructure score was 6% of optimal.

^aItems with 100% expert consensus on complete absence (n = 7), representing critical safety infrastructure.

^bSignificant difference between experts with =20 years vs. <20 years clinical experience (p = 0.004, Mann Whitney U test).

Examination bed scores were significantly higher (median = 2.5) than those of the seven critical safety items (all median = 0; p < 0.001, Mann–Whitney U), illustrating a “furniture–safety paradox” and highlighting AI’s tendency to prioritize visible furniture over essential safety infrastructure. Overall clinical usability was inadequate (median = 1.0), with composite scores indicating severe deficits in emergency preparedness (33%), critical monitoring infrastructure (6%), and testing infrastructure (13%). No differences were observed between experts with and without room design experience (all p > 0.05).

RoomReady’s photo-based assessment of two clinics revealed systematic strengths and limitations. In one clinic, the tool correctly identified visible safety features and flagged the lack of accessible epinephrine. However, it incorrectly reported a safety violation in the second clinic because the epinephrine was stored in a closed drawer. When prompted regarding a missing pulse oximeter, the tool accurately identified the gap and emphasized its clinical importance. This performance demonstrates RoomReady’s potential to detect the types of safety issues that motivated this study, despite its limitations in evaluating non-visible items. Expert validation using mDISCERN showed that RoomReady achieved a mean score of 36.44/40 (±2.50), corresponding to “good” quality. Inter-rater reliability was substantial (intraclass correlation coefficient = 0.81, 95% confidence interval 0.61–0.93), indicating consistent expert agreement regarding assessment quality across all four independent clinics (Table 2).

Table 2.

Expert validation of RoomReady using modified DISCERN (n = 16 evaluations)

mDISCERN item	Mean ± SD	Median	Mode	Range
1. Are the aims clear?	4.94 ± 0.25	5	5	4–5
2. Does it achieve its aims?	4.56 ± 0.51	5	5	4–5
3. Is it relevant?	4.63 ± 0.62	5	5	3–5
4. Are sources of information clear?	4.75 ± 0.45	5	5	4–5
5. Is timing of information clear?	4.50 ± 0.52	5	5	4–5
6. Is information balanced/unbiased?	4.44 ± 0.51	4	4	4–5
7. Additional support sources provided?	4.31 ± 0.70	4.5	5	3–5
8. Areas of uncertainty addressed?	4.31 ± 0.48	4	4	4–5
Total score (out of 40)	36.44 ± 2.50	37	38	31–39

SD: standard deviation.

Scores from 4 experts evaluating 4 independent clinics (16 total evaluations).

Inter-rater reliability: intraclass correlation coefficient = 0.81 (95% confidence interval, 0.61–0.93), indicating “Good” agreement.

Quality categories: Good (32–40), Fair (16–31), Poor (8–15).

RoomReady’s current limitations primarily reflect input constraints rather than algorithmic deficiencies: (1) items stored in closed locations remain undetectable without clear labeling; (2) assessment quality depends on photograph resolution, lighting, and viewing angles; and (3) equipment functionality cannot be verified from static images. Our findings demonstrate AI’s profound misunderstanding of medical space requirements, as the system prioritized aesthetic elements, such as clean surfaces and modern furniture, while omitting equipment essential for managing adverse reactions. The expert consensus regarding seven absent items reflects universally accepted safety standards rather than contested or discretionary equipment. While recent studies have shown AI improving healthcare efficiency by reducing patient wait times and falls [8,9], our findings reveal a dangerous naivety in AI’s understanding of medical safety priorities.

Conversely, the successful development and validation of RoomReady demonstrate AI’s potential when properly guided. By embedding comprehensive guidelines and expert knowledge into structured instructions, we transformed a system with dangerous blind spots into a practical quality improvement tool. RoomReady can identify critical gaps, such as missing pulse oximeters, that directly compromise patient safety. The allergy and immunology community increasingly recognizes AI as transformative, provided that implementation is accompanied by careful clinical oversight [ 10].

In conclusion, our study revealed critical blind spots in AI’s conceptualization of medical spaces while also demonstrating its potential as a safety assessment tool when trained using comprehensive international guidelines. Although developed for allergy test rooms, RoomReady’s methodology can be readily extended to other high-risk clinical environments by emphasizing safety-critical equipment and identifying gaps in facility readiness, including operating rooms, emergency departments, and intensive care units.