Research in Progress

ISH studies not only whether health technologies can work technically, but also what is required for them to be safely evaluated, integrated with human expertise, and used under real resource constraints. These programmes connect technical evaluation to questions about review capacity, supervision, infrastructure and workflow. Machine learning is a method within this work.

The studies below use retrospective data. They do not establish clinical usefulness or successful implementation in health services. Unpublished work can include completed experiments: the status of each workstream distinguishes experimental evidence, manuscript preparation and planned follow-up.

Selective Review for qPCR Interpretation

Research in progress — manuscript under author review

How should automated qPCR interpretation be combined with limited human review capacity so that difficult or uncertain reactions can be prioritised for expert attention?

This programme examines automated interpretation of amplification curves and selective referral for review. Retrospective evaluations across public qPCR data investigate confidence and uncertainty, disagreement with supplied reference labels, and how different referral strategies allocate a limited review workload. The operational question is which reactions remain automatically interpreted and which are prioritised for expert attention as review capacity changes.

Completed evaluations support current manuscript development and author review; this status does not mean that experiments are still running. Referral is evaluated as a routing decision. Correction of referred cases by human experts and actual time savings have not been demonstrated. Agreement with supplied amplification labels is not patient-level diagnostic performance. The work does not establish clinical diagnostic validation, prospective evaluation or deployment in Bangladesh.

Resource-Constrained Malaria Diagnostics

How can malaria image-analysis systems be evaluated and adapted for settings where annotation, computing resources, expert review and deployment infrastructure are constrained?

This programme brings together two related workstreams at different stages. Both concern research methods and their evaluation limits; neither provides a clinically validated diagnostic system.

Full-Field Malaria Detection under Limited Supervision

Manuscript in preparation

Follow-up evaluation planned

This work examines infected-cell localisation in full-field blood-smear images. It studies how annotation and supervision policies affect which targets contribute to training, and how training-data handling influences apparent performance. Evaluation considers false positives, missed targets and field-level performance, including the robustness of comparisons under limited supervision.

The original baseline is complete. Subsequent controlled experiments and annotation-retention audits provide completed developmental evidence for an exploratory, single-dataset manuscript. A bounded follow-up comparison has been designed, but no completed results from it are verified. It is optional to the current manuscript, and is not presented here as a running experiment or as completed evidence.

Repeated use of validation data limits performance interpretation. External-site validation has not been established, and qualified expert annotation review remains incomplete. The work does not establish patient diagnosis, parasitaemia estimation or clinical usefulness.

Edge-Efficient Malaria Image Classification

Experimental evaluation complete — manuscript in preparation

This work examines malaria classification in cropped blood-cell images under resource-constrained inference. Completed evaluations compare model compression and quantisation, CPU inference, sensitivity/specificity trade-offs and transfer to an external dataset. They address practical questions about computing requirements and whether a decision threshold selected on one dataset remains suitable on another.

The experimental campaign is complete and no additional model campaign is required. A manuscript has been prepared; author review and publication decisions remain unfinished. This is not ongoing model development or training.

Hardware measurements were made on a laptop CPU; phone and single-board-computer deployment were not evaluated. Decision thresholds transferred poorly across datasets, so successful threshold transfer is not claimed. The external evaluation does not establish independent-patient validation. These engineering comparisons do not demonstrate clinical usefulness or clinical deployment.

Evidence and implementation

The programmes identify questions for further evaluation, rather than evidence of service adoption or improved health outcomes. Connecting these methods to health-system practice would require appropriate independent evaluation, human oversight and assessment of local workflows, capacity and resources.

Status reviewed: 3 October 2026. Read about ISH's implementation-science and health-systems mission.