A study led by University College London (UCL) researchers found that implementing AI into NHS hospitals is more difficult than initially anticipated by healthcare leaders, with complications including around governance, contracts, data collection and staff training.

The study, published in The Lancet eClinicalMedicine on 10 September 2025,  examined a £21 million NHS England programme which launched in 2023 to introduce AI for the diagnosis of chest conditions, including lung cancer, across 66 NHS hospital trusts.

Researchers conducted interviews with hospital staff and AI suppliers to review how the diagnostic tools were procured and set up, and identify any pitfalls or factors that helped smooth the process.

They found that contracting took between four and 10 months longer than anticipated and by June 2025, 18 months after contracting was meant to be completed, a third (23 out of 66) of the hospital trusts were not yet using the tools in clinical practice.

First author Dr Angus Ramsey, principal research fellow at the UCL Department of Behavioural Sciences and Health, said: “Our study provides important lessons that should help strengthen future approaches to implementing AI in the NHS.

“We found it took longer to introduce the new AI tools in this programme than those leading the programme had expected.

“A key problem was that clinical staff were already very busy – finding time to go through the selection process was a challenge, as was supporting integration of AI with local IT systems and obtaining local governance approvals.

“Services that used dedicated project managers found their support very helpful in implementing changes, but only some services were able to do this.

“Also, a common issue was the novelty of AI, suggesting a need for more guidance and education on AI and its implementation.”

Challenges identified by the research included engaging clinical staff with high workloads in the project, embedding the technology in ageing and varied NHS IT systems across dozens of hospitals and a general lack of understanding and scepticism among staff about using AI in healthcare.

The researchers concluded that while “AI tools may offer valuable support for diagnostic services, they may not address current healthcare service pressures as straightforwardly as policymakers may hope”.

They recommend that NHS staff are trained in how AI can be used effectively and safely, and that dedicated project management is used to implement schemes like this in the future.

Senior author Professor Naomi Fulop at UCL, said: “The NHS is made up of hundreds of organisations with different clinical requirements and different IT systems and introducing any diagnostic tools that suit multiple hospitals is highly complex.”

The research, funded by the National Institute for Health and Care Research, was conducted by a team from UCL, the Nuffield Trust, and the University of Cambridge.

They are now studying the use of AI tools following early deployment when they have had a chance to become more embedded.

Researchers say that the findings should provide useful learnings on implementing the government’s 10 year health plan, published on 3 July 2025, which identifies AI as key to improving the NHS.

NYU research introduces video-based fire detection

The NYU Tandon School of Engineering has reported that its Fire Research Group has developed an artificial intelligence system that can detect fires and smoke in real time using existing CCTV cameras.

According to NYU Tandon, the system analyses video frames within 0.016 seconds, faster than a human blink, and provides immediate alerts.

The researchers explained that conventional smoke alarms activate only once smoke has reached a sensor, whereas video analysis can recognise fire at an earlier stage.

Lead researcher Prabodh Panindre, Research Associate Professor at NYU Tandon’s Department of Mechanical and Aerospace Engineering, said: “The key advantage is speed and coverage.

“A single camera can monitor a much larger area than traditional detectors, and we can spot fires in the initial stages before they generate enough smoke to trigger conventional systems.”

Ensemble AI approach improves accuracy

NYU Tandon explained that the system combines multiple AI models rather than relying on a single network.

It noted that this reduces the risk of false positives, such as mistaking a bright object for fire, and improves detection reliability across different environments.

The team reported that Scaled-YOLOv4 and EfficientDet models provided the best results, with detection accuracy rates above 78% and processing times under 0.02 seconds per frame.

By contrast, Faster-RCNN produced slower results and lower accuracy, making it less suitable for real-time IoT use.

Dataset covers all NFPA fire classes

According to the NYU researchers, the system was trained on a custom dataset of more than 7,500 annotated images covering all five fire classes defined by the National Fire Protection Association.

The dataset included Class A through K fires, with scenarios ranging from wildfires to cooking incidents.

This approach allowed the AI to generalise across different ignition types, smoke colours, and fire growth patterns.

The team explained that bounding box tracking across frames helped differentiate live flames from static fire-like objects, achieving 92.6% accuracy in reducing false alarms.

Professor Sunil Kumar of NYU Abu Dhabi said: “Real fires are dynamic, growing and changing shape.

“Our system tracks these changes over time, achieving 92.6% accuracy in eliminating false detections.”

Technical evaluation of detection models

NYU Tandon reported that it tested three leading object detection approaches: YOLO, EfficientDet and Faster-RCNN.

The group found that Scaled-YOLOv4 achieved the highest accuracy at 80.6% with an average detection time of 0.016 seconds per frame.

EfficientDet-D2 achieved 78.1% accuracy with a slightly slower response of 0.019 seconds per frame.

Faster-RCNN produced 67.8% accuracy and required 0.054 seconds per frame, making it less practical for high-throughput applications.

The researchers concluded that Scaled-YOLOv4 and EfficientDet-D2 offered the best balance of speed and reliability for real-world deployment.

Dataset preparation and training methods

The research team stated that it collected approximately 13,000 images, which were reduced to 7,545 after cleaning and annotation.

Each image was labelled with bounding boxes for fire and smoke, and the dataset was evenly distributed across the five NFPA fire classes.

The models were pre-trained on the Common Objects in Context dataset before being fine-tuned on the fire dataset for hundreds of training epochs.

The team confirmed that anchor box calibration and hyperparameter tuning further improved YOLO model accuracy.

They reported that Scaled-YOLOv4 with custom training configurations provided the best results for dynamic fire detection.

IoT cloud-based deployment

The researchers outlined that the system operates in a three-layer Internet of Things architecture.

CCTV cameras stream raw video to cloud servers where AI models analyse frames, confirm detections and send alerts.

Detection results trigger email and text notifications, including short video clips, using Amazon Web Services tools.

The group reported that the system processes frames in 0.022 seconds on average when both models confirm a fire or smoke event.

This design, they said, allows the system to run on existing “dumb” CCTV cameras without requiring new hardware.

Deployment framework and false alarm reduction

The NYU team explained that fire detections are validated only when both AI models agree and the bounding box area grows over time.

This approach distinguishes real flames from static images of fire, preventing common sources of false alerts.

The deployment is based on Amazon Web Services with EC2 instances handling video ingestion and GPU-based inference.

Results and metadata are stored in S3 buckets and notifications are sent through AWS SNS and SES channels.

The researchers stated that this cloud-based framework ensures scalability and consistency across multiple camera networks.

Applications in firefighting and wildland response

NYU Tandon stated that the technology could be integrated into firefighting equipment, such as helmet-mounted cameras, vehicle cameras and autonomous robots.

It added that drones equipped with the system could provide 360-degree views during incidents, assisting fire services in locating fires in high-rise buildings or remote areas.

Capt. John Ceriello of the Fire Department of New York City said: “It can remotely assist us in confirming the location of the fire and possibility of trapped occupants.”

The researchers noted that the system could also support early wildfire detection, giving incident commanders more time to organise resources and evacuations.

Broader safety applications

Beyond fire detection, the NYU group explained that the same AI framework could be adapted for other safety scenarios, including medical emergencies and security threats.

It reported that the ensemble detection and IoT architecture provide a model for monitoring and alerting in multiple risk environments.

Relevance for fire and safety professionals

For fire and rescue services, the system demonstrates how existing CCTV infrastructure can be adapted for early fire detection without requiring new sensors.

For building managers, the research shows how AI video analysis could supplement or back up smoke alarms, particularly in settings where detector failure is a risk.

For wildland and urban response teams, the ability to embed the system into drones or helmet cameras may improve situational awareness and decision-making during fast-developing incidents.

AI system uses CCTV to detect fires in real time: Summary

The NYU Tandon School of Engineering Fire Research Group has reported an AI system that detects fires using CCTV cameras.

The research was published in the IEEE Internet of Things Journal.

The system processes video at 0.016 seconds per frame.

Scaled-YOLOv4 achieved 80.6% accuracy and EfficientDet achieved 78.1% accuracy.

False detections were reduced by tracking bounding box changes over time.

The dataset included 7,545 images covering all five NFPA fire classes.

Alerts are generated in real time through AWS cloud systems.

Applications include CCTV monitoring, drones, firefighter equipment and wildland detection.

The research suggests the same framework could support wider emergency monitoring.

AI is racing ahead. Regulation? Not so much. Kevin Williamson talks with Adam Thierer, senior fellow at the R Street Institute, about the opportunities and risks of artificial intelligence. They dive into the policy fights shaping its future, the role of Big Tech, and how AI could reshape global competition.

The Agenda:
—Defining AI
—Hardware vs. software
—Economic and geopolitical implications of AI
—Job replacement concerns
—Tech skeptics

The Dispatch Podcast is a production of The Dispatch, a digital media company covering politics, policy, and culture from a non-partisan, conservative perspective. To access all of The Dispatch’s offerings—including access to all of our articles, members-only newsletters, and bonus podcast episodes—click here. If you’d like to remove all ads from your podcast experience, consider becoming a premium Dispatch member by clicking here.