Improving tuberculosis diagnosis in Kenyan public laboratories with an affordable offline **AI **assistant

In Kenya, microscopy remains the most common method of diagnosis for tuberculosis and malaria, and is often the only option available in remote areas (MoH, 2020). However, the reliability of this technique is highly dependent on the healthcare professional's skills, workload, and the condition of equipment, particularly in peripheral laboratories, already under pressure and with scant resources, where the majority of diagnoses are made (Carter, 2016).

The limitations of microscopy in tuberculosis diagnosis have been well documented (Wongsrichanalai et al., 2007; Obare et al., 2013). In smear microscopy for tuberculosis, sensitivity varies from 20% to 80% depending on the reader's expertise, the quality of staining, and the bacillary load (Steingart et al., 2014).

This variability has direct repercussions for patients. A false negative delays life-saving care and prolongs the period of contagion, while a false positive exposes patients to unnecessary treatment and means that alternative diagnoses are missed. These errors result in preventable deaths and increase the financial burden on low-income families.

This project aims to provide a low-cost, fully offline, AI-powered device, used in complement with microscopes, to increase diagnosis accuracy in Kenyan public health laboratories where resources are scarce.

This innovative system offers several features that set it apart from existing solutions:

• Fully offline operation, avoiding the need for cloud infrastructure or an internet connection. The algorithm uses compression techniques adapted to low-power devices.
• Compatibility with microscopes already used in Kenyan public laboratories. Only minor modifications are needed to fit the adapter to existing microscopes, and the system is supplied by a rechargeable battery that continues to operate even during power outages.
• A model trained on local data, using annotated sputum slides collected in Kenya. The algorithm is calibrated to detect acid-fast bacilli under real-world conditions of staining, smear quality, and optical variation, specific to Kenya.

Once assembled, these components transform a standard microscope into an AI-powered microscope. No data transmission is required: the analysis is performed directly on a low-power device. During the preparation phase, the prototype will be tested and fine-tuned in ten pilot laboratories across six municipalities (Siaya, Busia, Kisumu, Vihiga, Homa Bay, and Migori), selected to cover the broad range of tuberculosis diagnosis conditions in Kenya.

Following the preparation phase, the consortium aims to demonstrate the solution's technical feasibility, acceptability, and potential deployment in Kenyan public laboratories. The project has three deliverables:

• Produce a validated prototype, tested under real-world conditions in ten pilot laboratories across six municipalities, that operates on low-power devices and can be adapted to existing workflows without the need to modify existing equipment;
• Conduct an acceptability evaluation, with the 45 to 55 people directly involved in the preparation phase – laboratory technicians, supervisors, and technical staff from the consortium – through co-design workshops and user testing.