Healthcare & Life Sciences Cloud Solutions

Compliance is the architecture, not a review at the end

Health data is the one category where getting it wrong is unrecoverable. You can rotate a leaked API key. You can’t un-disclose someone’s diagnosis, or their genome. Sequencing data is worse than most people assume: it identifies a patient for life, it can’t be reissued, and it implicates relatives who never consented to anything.

So we design the controls first. Patient workloads are pinned to the regions you nominate, Sydney and Melbourne for Australian providers, Singapore for those operating under MOH oversight, with customer-managed encryption keys so Google holds the ciphertext and you hold the key. Organisation policy constraints stop a well-meaning engineer spinning up a resource in the wrong country at all. VPC Service Controls draw a perimeter that data can’t be copied out of, even by someone with valid credentials. De-identification happens on ingest through the Cloud Healthcare API’s de-id pipelines and Cloud DLP, so the analytics copy never contains a national health identifier in the first place. That ordering matters: masking after landing gives you a clean table and a dirty bucket behind it.

Which obligations apply depends on where you operate. In Australia that’s the Privacy Act and the Australian Privacy Principles, the My Health Records Act where a provider connects to it, and the Notifiable Data Breaches scheme, which is what makes the audit trail load-bearing rather than nice to have. In Singapore it’s the PDPA, plus the Healthcare Services Act licensing conditions and the data handling rules that come with connecting to the National Electronic Health Record. For groups with US or EU arms we add HIPAA and GDPR boundaries, normally as separate projects rather than separate tags, because tags don’t stop a query.

Multi-jurisdiction is where this gets genuinely hard, and it’s worth being blunt about it. A single regional data platform serving both Australian and Singaporean patients usually can’t be done as one dataset with a country column. Cross-border disclosure rules mean the practical answer is separate projects per jurisdiction with a federated query layer above them, which costs more and is slower to build than the architecture people usually arrive asking for. Research collaborations then add ethics-committee conditions on top, and those are often stricter than either country’s legislation.

Genomics: the compute problem is bursty, not big

A genomics pipeline is idle most of the month and then wants a thousand cores for six hours. Buying for the peak wastes most of the year, and queuing on a shared cluster means researchers wait weeks.

We run these on Google Batch with Nextflow or Cromwell, so the cluster exists for the duration of the run and then stops. Variant calling uses DeepVariant, Google’s deep-learning caller, which reads the pileup as an image and is measurably more accurate on indels than the older statistical callers. DeepSomatic covers tumour-normal work. Called variants land in BigQuery rather than a pile of VCFs on a share, which is what makes cohort queries across thousands of samples take seconds instead of a batch job. For structural biology, AlphaFold runs on Vertex AI against the same storage.

None of that is a language model. It’s convolutional and transformer-based inference over sequence and image data, which is where the accuracy in this field actually comes from.

We’ve built this shape of burst compute before, including an on-demand 1,000-core cluster for a global engineering client. The scheduling problem is the same whichever domain the workload comes from.

Predicting population health at scale

This is the part most health organisations underuse, and it doesn’t need generative AI at all. Predicting emergency department demand, hospital readmission risk, chronic disease progression or immunisation coverage gaps is tabular and time-series forecasting. The right tools are boosted trees and sequence models, not chat.

BigQuery ML is usually where we start, because the data is already there and moving it is the expensive part. ARIMA_PLUS handles seasonal demand forecasting with holiday effects, which matters when your peak follows public holidays and school terms rather than a smooth curve. Boosted-tree classifiers handle risk stratification, and they have the advantage of explaining themselves: feature importance is something a clinical governance committee can interrogate. Where a model needs to be custom, Vertex AI Forecast and custom TensorFlow training run against the same de-identified tables.

A caveat worth stating. These models learn from historical care patterns, so they inherit historical access inequities. A readmission model trained on a population that under-presented to hospital will under-predict for that population. We build cohort-level bias checks into the evaluation step, and we’d rather flag a model as unsafe to deploy than ship it with a good aggregate AUC. Sometimes the honest answer is that the data isn’t good enough yet.

Where generative AI does earn its place

Narrow, human-in-the-loop, back-office work. Turning a dictated note into structured fields, suggesting clinical codes for a coder to accept or reject, triaging an intake form so the right person reads it first. Document AI and Vertex AI handle these well, with a person signing off every output. We don’t build anything that puts a model between a clinician and a diagnosis.