Why Science? | BloodCounts! Foundation Model of Blood

The complete blood count, also called full blood count in Singapore and the UK, is widely used to measure blood-cell features. The project asks whether raw analyser information usually left outside the standard report can support new research. It uses an existing sample and measurement stream, but additional analysis still needs governance and validation.

A routine full blood count contains familiar summary values and much richer raw instrument data; research can ask whether discarded signals carry reproducible information without turning every pattern into a diagnosis. The chapter develops laboratory-test literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Separate the standard report, raw instrument data and later model output in a diagram. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Extra data do not automatically create clinically useful information.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Next chapter

NRF names Cambridge CARES and NTU as co-leads. CARES’ October 2026 release identifies Professor Weisi Lin and Professor Michael Roberts as leads, with NUHS and SingHealth as local hospital partners. Roles, datasets and outputs should be taken from current official records.

A foundation model of blood can learn across many samples and later tasks, but clinical usefulness depends on diverse populations, calibration, intended use, prospective validation and regulatory review. The chapter develops programme literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Map research, clinical, data and engineering responsibilities among partners. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A partner list does not imply a deployed diagnostic tool.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

An FBC reports measures related to red cells, white cells, haemoglobin and platelets, helping clinicians investigate conditions such as anaemia or infection in context. Results depend on the patient, laboratory ranges, history and other tests. The project does not replace clinical interpretation of the current report.

BloodCounts! combines numerical measurements with blood-smear images, joining laboratory medicine, imaging, mathematics and clinical context inside one carefully governed evidence system. The chapter develops clinical-laboratory literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Build a sample report map showing measurement, reference context and clinician question. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Only qualified professionals can interpret an individual result.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Automated instruments record distributions, scatter patterns and other signals while classifying cells. Standard workflows report selected parameters, while richer raw information may be discarded. Research can test whether these signals are reproducible across machines and associated with outcomes beyond known measures.

Federated learning can support collaboration while data remain within trusted environments, yet privacy, governance, harmonisation and site-specific bias still require active management. The chapter develops measurement provenance. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create a provenance map from instrument event to reported and retained fields. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A hidden instrument pattern is not automatically a disease marker.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

A blood smear places cells on a slide for microscopic review. NRF notes that a proportion of FBCs at SingHealth are supplemented with smear examination. Images can show morphology that numerical summaries miss, but preparation, staining, scanning and expert labels all introduce variation.

A routine full blood count contains familiar summary values and much richer raw instrument data; research can ask whether discarded signals carry reproducible information without turning every pattern into a diagnosis. The chapter develops medical-imaging data literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Write an image-data card with preparation, scanner, annotation and quality checks. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: An image model cannot substitute for validated laboratory and clinical workflows.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

CARES describes a large model trained on full raw CBC/FBC parameters, supplemented by smear images. Pretraining across many samples may support different later tasks. Each application needs separate evaluation because a representation useful for one condition may be miscalibrated or irrelevant for another.

A foundation model of blood can learn across many samples and later tasks, but clinical usefulness depends on diverse populations, calibration, intended use, prospective validation and regulatory review. The chapter develops foundation-model evaluation. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Define pretraining evidence and two downstream validation plans. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Foundation-model status does not confer clinical approval.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The project says the model will be trained without optimisation for one condition, aiming for adaptability. This can reveal broad patterns but also discover correlations unrelated to disease mechanisms. Later studies must define the clinical question, population, outcome and comparator before claiming usefulness.

BloodCounts! combines numerical measurements with blood-smear images, joining laboratory medicine, imaging, mathematics and clinical context inside one carefully governed evidence system. The chapter develops intended-use specification. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Turn an open-ended pattern into a precise intended-use study question. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Exploratory discovery is not a validated screening programme.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Not every test has a smear image, and smears may be ordered because something already looked unusual. Combining numerical and visual data must represent this selection process so the model does not treat image availability as a shortcut. Compare modes alone and together, and audit missing-data patterns.

Federated learning can support collaboration while data remain within trusted environments, yet privacy, governance, harmonisation and site-specific bias still require active management. The chapter develops multimodal-model auditing. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create a modality-availability matrix and test for ordering bias. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Missing modalities are not random by default.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Different analysers, laboratory procedures, coding systems and patient populations can shift data. CARES describes curation in trusted hospital environments and return of models for local analysis. Harmonisation should preserve site labels, instrument versions and quality controls so performance can be calibrated rather than averaged blindly.

A routine full blood count contains familiar summary values and much richer raw instrument data; research can ask whether discarded signals carry reproducible information without turning every pattern into a diagnosis. The chapter develops multisite validation. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Build a site-and-instrument stratified validation table. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Pooling data does not erase hospital-specific variation.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The international BloodCounts! Consortium uses a federated approach in which models can be trained or evaluated across institutions without centralising all raw data. This reduces some transfer risk, but model updates can still leak information and sites need compatible definitions, security and governance.

A foundation model of blood can learn across many samples and later tasks, but clinical usefulness depends on diverse populations, calibration, intended use, prospective validation and regulatory review. The chapter develops privacy-preserving collaboration. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Draw a federated workflow with local data, shared updates and audit controls. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Federation is not automatic privacy or equivalence.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

CARES lists stroke as an initial application. A study should define whether the goal is risk prediction, acute support, prognosis or another question, along with the timing of the blood test and clinical comparator. Retrospective signals need prospective evaluation before they can change care.

BloodCounts! combines numerical measurements with blood-smear images, joining laboratory medicine, imaging, mathematics and clinical context inside one carefully governed evidence system. The chapter develops clinical-study design. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Write a stroke-study intended-use statement with time horizon and decision. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A statistical association does not diagnose or predict an individual stroke.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The project also names lung-cancer studies. Evaluation should specify population, disease stage, reference standard and what action a model result would trigger. Sensitivity, specificity, calibration, false positives and missed cases carry different consequences. Compare with current pathways rather than testing the model in isolation.

Federated learning can support collaboration while data remain within trusted environments, yet privacy, governance, harmonisation and site-specific bias still require active management. The chapter develops screening-evidence literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Build a screening-evidence table with benefits, harms and follow-up. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Research signals do not establish a recommended cancer-screening test.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Open and federated models may help rare-disease research where one hospital has few examples. Small samples still produce wide uncertainty, and phenotype definitions may differ. Researchers should report confidence intervals, external validation and whether the model genuinely adds evidence rather than memorising a few cases.

A routine full blood count contains familiar summary values and much richer raw instrument data; research can ask whether discarded signals carry reproducible information without turning every pattern into a diagnosis. The chapter develops rare-disease methods literacy. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Design a rare-condition evaluation with uncertainty and leave-one-site-out testing. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: International sample aggregation does not guarantee representativeness.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

CARES says Singapore partners add population diversity to an international consortium. Fair evaluation should examine age, sex, ancestry, clinical setting, analyser and relevant health conditions, with sample sizes and calibration reported. Differences require investigation rather than automatic correction or biological storytelling.

A foundation model of blood can learn across many samples and later tasks, but clinical usefulness depends on diverse populations, calibration, intended use, prospective validation and regulatory review. The chapter develops health-AI fairness auditing. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create a subgroup audit with measurement and selection explanations. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: A diverse dataset can still contain underrepresented groups and biased labels.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

A useful model must deliver information at the right time, integrate with laboratory and clinical systems and improve a decision without creating excessive alerts or delay. Prospective studies, human-factors work and post-deployment monitoring examine performance in the real pathway. Clinicians remain accountable for patient care.

BloodCounts! combines numerical measurements with blood-smear images, joining laboratory medicine, imaging, mathematics and clinical context inside one carefully governed evidence system. The chapter develops clinical-workflow evaluation. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Map model output to reviewer, action, fallback and monitoring. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Retrospective model accuracy is not demonstrated patient benefit.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

The project combines mathematics, computer science, bioinformatics, imaging, haematology, oncology, neurology and clinical informatics. Students can learn from de-identified teaching datasets and focus on calibration, uncertainty and ethics. Current institutions control qualifications, data access and professional practice.

Federated learning can support collaboration while data remain within trusted environments, yet privacy, governance, harmonisation and site-specific bias still require active management. The chapter develops health-data career preparation. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Plan a safe teaching analysis comparing a simple baseline with a calibrated model. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Educational analysis provides no basis for interpreting personal blood tests.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter · Next chapter

Archive the official project edition, hospital and analyser, consent and governance basis, data dictionary, smear workflow, model version, training scheme, intended use, validation, subgroup results, human review, regulatory status and monitored outcomes. This separates research promise from clinical evidence.

A routine full blood count contains familiar summary values and much richer raw instrument data; research can ask whether discarded signals carry reproducible information without turning every pattern into a diagnosis. The chapter develops evidence governance. Keep the official scope, input version, method, raw evidence, uncertainty, review and next decision together. That receipt shows what the model suggested, what the real-world check found and where the claim must stop. NRF and CARES describe the project; NUHS, SingHealth, NTU, Cambridge, data-governance bodies, clinicians and regulators control data access, research use, clinical decisions and deployment.

Put this into practice by turning the idea into a small, reviewable action: Create a versioned test–data–model–study–decision lineage. Then ask what the resulting note or artefact actually demonstrates. A useful record names context, source or instruction, action, observation, interpretation and uncertainty. It also preserves this limit: Current hospitals, ethics bodies and regulators control clinical use.

An evidence check keeps confidence proportional to what the work can show. Separate the original observation or official statement from the learner’s explanation. Record feedback and any revision. If information is missing, write “not shown” or “not yet known” rather than guessing. That honest gap can become the next question.

Contents · Previous chapter