Systems that look for the patients the system loses

The problem

The starting point combined long delays in specialist access, critical lags in diagnostic confirmation and systematic breaches of the legal deadlines set by the GES regime. Paper referrals stalled at administrative handoffs, and nobody knew where each patient actually was.

Architecture

1. IntakeCancer suspicion referred from primary care or the emergency department.
2. FlaggingA fluorescent sticker on the chart and the paper referral: the flag travels with the document through every administrative handoff.
3. RegistrationEntry into the traceability sheet with a named clinical case manager assigned.
4. Follow-upActive contact at each milestone: first appointment, diagnostic confirmation, tumour board and treatment decision.
5. EscalationAlert to hospital management when a GES deadline is at risk, before it expires.
6. ExitTreatment, or early referral to palliative care and pain relief.

What it produced

• More than 6,000 patients entered the traceability system over the period.
• Time to first specialist appointment brought down to 13 days.
• The hospital placed among the five best public hospitals in Chile for cancer management.
• Sustained stretches with zero overdue GES guarantees in neoplastic conditions.
• Faster referral to palliative care and pain relief for patients at advanced stages.

The problem

Patients with suspected or evident cancer sit in the waiting list classified as general or non-prioritised elective surgery, losing the priority they are entitled to. The system covers, by design, cancers without an explicit guarantee — melanoma, sarcomas, head and neck, oesophagus, pancreas, hepatobiliary, central nervous system, endometrium — because the sixteen GES cancers already have their own guaranteed pathway. Those are precisely the ones nobody was counting.

Architecture

1. ExtractionA cut of the national surgical waiting list from the layered warehouse into Parquet, queried with DuckDB: ~475,000 diagnoses.
2. Deterministic filteringClinical regular expressions and ontological mapping over ICD-10 capture unambiguous criteria before any model is involved.
3. VectorisationEmbeddings of the free text against centroids of confirmed cancer diagnoses. The normalised text’s hash is the cache key, so only new text is vectorised.
4. ClassificationA gradient-boosted tree ensemble; the threshold is recalibrated with the Youden index on every retraining.
5. DecisionThree states. Suspicious and indeterminate are mandatorily routed to medical clinical audit; the model closes nothing.
6. DeliveryAn official, auditable monthly cut, plus a weekly preliminary channel over the transactional base surfacing cases with about 7 days of lag.

What it produced

• Classifies roughly 475,000 diagnoses from the national surgical waiting list on every cut.
• Evaluated on 15,032 cases individually validated by clinical specialists: 91.6 % sensitivity, 97.9 % specificity, 82.8 % precision (PPV) and 99.0 % negative predictive value.
• The label is official and auditable: it is what the care network uses to prioritise cases with cancer indications.
• Monthly validation against the Health Services’ own criteria, with nine reports per service, waiting times by cancer stage and the indicator for cancer surgery within 90 days.
• The loop does not end at classification: weekly comparison against each service’s validation surfaces miscoded records and drives their correction at source, with the clinical leads of each service.

Evaluation

The problem

Decisions on surgical capacity and budget allocation are made looking at the current stock of the list, which only describes the past. Projecting requires modelling entries, exits and their seasonality separately, and those dynamics differ too much between territories for a single national model to be useful.

Architecture

1. InputHistorical series of monthly cuts, broken down by Health Service.
2. ModelEntries and exits are modelled separately, each with its own seasonality: the stock is the result, not the variable.
3. CalibrationOne fit per Health Service instead of a single national model, because territorial dynamics are not comparable.
4. VerificationBacktesting over historical cuts and a 30-check adversarial audit designed to break the model.
5. OutputAn in-browser simulator where the decision-maker moves capacity and allocation-policy assumptions.

What it produced

• Moves the conversation from the current stock to the list’s expected trajectory.
• The simulator puts the model directly in the hands of the decision-maker, which is where the assumptions actually get argued.

The problem

Multiple identifying keys distort the care database. Migrants awaiting visa processing, newborns and undocumented people enter under temporary numeric identifiers and later receive their definitive national ID, with no automatic system-side merge of their care events.

Architecture

1. Source2.1 million records with multiple identifying keys per person.
2. EngineSplink over DuckDB, in-process: blocked comparison runs on the same machine, with no cluster.
3. ComparisonJaro-Winkler similarity over name, date of birth, sex and address, weighted under the Fellegi-Sunter framework.
4. Model selection40 candidate configurations are trained and the one that separates matches from non-matches most cleanly is kept, measured with Hartigan’s dip test over the score distribution.
5. OutputUnified identities that correct the count of real surgical demand.

What it produced

• Corrects the overstatement of surgical waiting lists caused by redundant records.
• Objective indicators of real demand, underpinning sector budget allocation and Ministry of Health policy.