SPECTRA AI-TB: Smart Predictive Evaluation and Chest X-ray Artificial Intelligence Analyzer for Tuberculosis in Kuala Nerus

The Problem

​Tuberculosis (TB) remains a major infectious disease challenge requiring early detection and rapid clinical intervention to stop community transmission. However, conducting active screening in primary healthcare settings and district outreach programs without on-site stationary X-ray facilities or resident radiologists presents severe logistical barriers.

​Traditionally, patients in underserved coastal and island communities had to travel across district borders to secondary referral hospitals just for a routine chest radiograph (CXR). Standard radiological reports often took days to weeks to return. This reporting delay led to high rates of patients being lost to follow-up before microbiological testing or treatment could begin, created significant administrative workloads, and prolonged the diagnostic cascade.

​The Tuberculosis/Leprosy Unit at the Kuala Nerus District Health Office sought an objective, rapid screening tool to streamline front-line triage while preserving diagnostic accuracy and clinical safety.

​The Solution

​The team developed SPECTRA AI-TB (Smart Predictive Evaluation and Chest X-ray Artificial Intelligence Analyzer for Tuberculosis), an AI-driven clinical decision support system engineered to evaluate digital chest radiographs and produce rapid, structured triage reports at the point of care.

​How it works

Healthcare personnel upload digital DICOM or standardized CXR images captured from portable or clinic X-ray units into the SPECTRA AI-TB interface:

1. ​Deep convolutional neural networks evaluate the lung fields for radiological patterns suggestive of active pulmonary TB (such as apical infiltrates, cavitation, consolidation, and miliary patterns).
2. ​The model outputs a quantitative risk score (0–100) benchmarked against an optimal clinical decision threshold of \tau = 50.
3. ​The platform automatically compiles a structured, bilingual radiology report draft (in Bahasa Melayu and English) detailing pulmonary findings within an average runtime of 2.30 minutes.

​How it fits into the clinical workflow

SPECTRA AI-TB operates strictly as an assistive "second reader" and triage prioritization tool rather than an autonomous diagnostic system:

• ​Medical Officer Oversight: The attending Medical Officer evaluates the patient, reviews the original radiograph alongside the AI-generated assessment, and makes the definitive decision to order sputum GeneXpert/AFB confirmation or initiate clinical management.
• ​Traceable Audit Chain: The system operates under a strict data governance framework: DATA \rightarrow CALCULATION \rightarrow RESULT \rightarrow FIGURE. Every score, threshold, and classification decision is fully reproducible and auditable, eliminating black-box uncertainty in clinical triage.
• ​Same-Day Community Action: During Active Case Detection (ACD) field operations, immediate high-risk alerts allow the team to collect on-the-spot sputum samples from presumptive cases before individuals leave the screening venue.

​Patient awareness and privacy

The system complies with health data privacy standards by processing de-identified imaging data without requiring sensitive identity markers (such as national ID numbers) for model inference. Community members participating in mobile screenings are briefed transparently on the use of AI-assisted triage, retaining the right to standard manual review by medical officers.

​The Impact

​The deployment and scientific verification of SPECTRA AI-TB have demonstrated substantial clinical, operational, and financial gains:

​1. Measurable operational turnaround

• ​Instantaneous Reporting: CXR analysis and report drafting time was cut from several days to an average of 2.30 minutes per case, enabling immediate, same-day clinical decision-making.
• ​Elimination of Dropouts: Near-instant triage substantially mitigated loss-to-follow-up across remote communities by securing presumptive TB cases immediately for sputum collection.

​2. Independent validation results

A controlled internal validation cohort of 200 chest radiographs benchmarked against independent expert reference standards demonstrated:

• ​Sensitivity: 94.2% (113/120 confirmed positive cases detected).
• ​Specificity: 91.3% (73/80 non-TB cases cleared).
• ​Diagnostic Accuracy: 93.0% overall accuracy with a peak Youden Index of 0.854 at \tau = 50.
• ​Discriminative Power: An exact Receiver Operating Characteristic Area Under the Curve (ROC AUC) of 0.9600 (95% CI: 0.934–0.986).
• ​High Usability and Satisfaction: The System Usability Scale (SUS) assessment yielded a mean score of 85.5 (Grade A) among operators. Clinical satisfaction reached 93.75% among Medical Officers and 96.40% among community participants.

​3. Economic and health system savings

Operational cost modeling demonstrated net operational expenditure (OPEX) savings of RM90,050 per year (a 48.2% reduction) by eliminating redundant patient transport, minimizing cross-district travel subsidies, and replacing paper-heavy manual reporting overheads.

​Lessons Learned

• ​End-to-end data auditability is vital for clinical AI adoption. Clinical trust does not stem from headline accuracy metrics alone, but from whether every diagnostic figure and threshold can be mathematically mapped back to a transparent master dataset.
• ​AI delivers peak public-health value at the geographic periphery. Deploying CDSS tools in settings without fixed imaging facilities democratizes specialist-level radiological triage, bringing hospital-tier screening capabilities directly to primary healthcare and outreach settings.
• ​A human-in-the-loop architecture preserves safety and workflow buy-in. Presenting AI output as a structured draft report for clinician sign-off rather than an automated verdict removes clinician hesitation, satisfies regulatory requirements, and aligns with established healthcare governance.

​Regulatory Context

​The platform was designed to align with Malaysia's healthcare regulatory landscape and international guidelines for Software as a Medical Device (SaMD). Data workflows adhere to the Personal Data Protection Act (PDPA) by stripping personal identifiers prior to inference. All data assets and inference services are managed locally to ensure data sovereignty and clinical accountability within the public health network.

​Related Projects & Resources

• ​NIH Chest X-ray dataset

Registered under the National Medical Research Register (NMRR Research ID: RSCH ID-26-04812-ANU) and preparing for Software as a Medical Device (SaMD) registration with the Medical Device Authority (MDA) Malaysia.