AI in Malaysia · Healthcare

What 14,350 chest X-rays say about AI in Malaysian healthcare

The Ministry of Health's AI-assisted lung screening found something its designers did not lead with. The lesson applies to every clinic and healthcare company thinking about AI: screen broadly, and keep a person in the decision.

Pikabyte Engineering 3 min read

25.5%

of 14,350 AI-assisted chest X-rays at seven clinics showed abnormal findings

In 2025 the Ministry of Health began using AI-assisted chest X-ray screening at seven primary care clinics under its Lung Health Initiative. A circular reported by CodeBlue later that year gave the early numbers: 14,350 people screened, with 25.5% recording abnormal findings.

The headline case for the programme was lung cancer. An earlier pilot had found 22 confirmed lung cancer cases among 3,482 X-ray examinations over two months. But the early rollout detected far more mild, asymptomatic tuberculosis than cancer. The model flagged what was on the film, and what was on the film was not quite what the programme had been named for.

In May 2026 the Health Minister said the ministry was targeting 902 clinics for the integrated lung health rollout by the end of the year.

Three lessons from the rollout

AI widens the net. It does not make the diagnosis. The screening model flags a film as abnormal. Spirometry, TB testing and low-dose CT for selected high-risk patients follow, and a clinician decides. This is the correct shape for AI in any high-stakes setting: the model sorts, people decide.

Real deployments surprise you. A system designed around one condition surfaced another at higher volume. That is only useful if the workflow downstream can absorb it. An AI project that plans for the model and not for the queue it creates will stall in its first month.

Scale is an operations problem. Going from seven clinics to several hundred is not about a better model. It is about equipment, connectivity, training and follow-up capacity.

The quieter opportunity: the paperwork around care

Clinical AI gets the attention and the regulatory scrutiny, rightly. For most clinics, hospitals and healthcare companies, the faster win is administrative.

Healthcare runs on documents in inconsistent formats. Referral letters, rosters, job postings, claim forms, lab reports, all arriving as PDFs, spreadsheets and photographs of paper, all needing to be typed into a system by someone who has better things to do.

We built an intake pipeline for a medical staffing platform in Japan. Institutions sent part-time job postings as PDFs, PowerPoint slides, spreadsheets and photos. Staff transcribed each one. The pipeline now accepts any of those formats, uses a fine-tuned model to extract the role into a strict schema, and creates a record that is queued for a person’s review before it is published.

Manual data entry and review time fell by 76%. It went live in three and a half weeks. Nothing about it touches a clinical decision, which is exactly why it could move that fast.

A sensible order of work

  1. Start with intake. Any process where staff retype information from a document is a candidate. The risk is low because a person confirms the result, and the saving is immediate.
  2. Constrain the output. The model should fill a fixed schema, and anything that does not fit should be rejected, not guessed.
  3. Queue for review. Every AI-created record waits for a person before it affects a patient, a clinician or a payment.
  4. Treat the data as sensitive. Health information is sensitive personal data. Know which provider processes it and where, and be ready to tell a patient.
  5. Leave clinical AI to validated products and the clinicians accountable for using them.

The ministry’s screening programme and a staffing company’s intake pipeline look nothing alike. They are built on the same idea. Let the machine read everything, and let a person make the call.

Sources

Written by

Pikabyte Engineering

Pikabyte builds custom business systems, applied AI and secure cloud infrastructure from Malaysia. Talk to us.

The work behind this article

Case study · Healthcare · Medical staffing

Unstructured job postings in any format, structured listings out.

An AI pipeline that turns part-time medical job postings from PDFs, slides, spreadsheets and photographed documents into structured, review-ready records for a healthcare staffing platform in Japan.

76%
less manual data entry and review time
23+
job postings extracted per day on average
3.5 weeks
from kickoff to production
4
input formats: PDF, slides, sheets, photos
Read the case study On Call Inc. · Japan

Start a conversation

Want this applied to your business?

Tell us what is slowing the business down. You will talk to an engineer, not a salesperson, and get a straight answer on whether software is the right fix.