During employee medical check-up (MCU) cycles, examining physicians evaluate extensive multi-system findings: blood panels, spirometry curves, audiograms, ECG recordings, and physical findings. The clinical outcome must determine whether an employee is fit, fit with restrictions, or unfit for specific job hazards.
With emerging clinical software, questions arise regarding whether algorithmic scoring can automate this determination.
At the 1st PERDOKI Jakarta Summit (24 October 2026), an AI-assisted focused review screened 1,307 unique scientific records examining a core question: how accurately do automated scoring tools and algorithms identify workers that qualified physicians judge unfit?
The review demonstrates that empirical diagnostic accuracy data remains scarce, highlighting why independent physician judgment remains essential for worker safety.
The Literature Landscape: Scarcity of Paired Evidence
The review evaluated literature across PubMed, Europe PMC, and regional Indonesian clinical indices using an unregistered, post-hoc protocol.
Of 2,492 initial records, 1,307 unique studies underwent title and abstract screening:
- 43 full-text articles were sought for detailed appraisal.
- 4 studies (across 5 published reports) provided quantitative 2x2 data comparing scoring criteria against physician determinations.
- Within this review, validated paired counts were identified in only two occupational cohorts: Indonesian oil-and-gas workers and municipal hazardous-materials (HAZMAT) firefighters.
Across other employment sectors screened in this review, peer-reviewed diagnostic accuracy studies comparing automated scoring against independent physician panels were not identified.
Examining the Data: Sensitivity and Specificity
In occupational fitness assessment, missed pathology (a false negative where an unfit worker is categorized as fit) presents substantial workplace safety risks. In diagnostic accuracy terms, unfit serves as the target condition.
The identified cohorts demonstrated distinct performance metrics:
| Occupational Cohort | Evaluated Criteria | Sensitivity for Unfit (95% CI) | Specificity for Fit (95% CI) | Clinical Context |
|---|---|---|---|---|
| Indonesian Oil & Gas Personnel (n=607) | Multi-parameter weighted score | 79.17% (57.85% – 92.87%) | 98.80% (97.54% – 99.52%) | Missed 5 of 24 specialist-judged unfit examinations. Cut-off was optimized on the same cohort without external validation. |
| Municipal HAZMAT Firefighters (n=340) | Draft NFPA 1582 medical criteria | 11.11% (0.28% – 48.25%) | 98.79% (96.93% – 99.67%) | Agreed with 1 of 9 physician unfit determinations; criteria missed 8 of 9 unfit cases. |
| Municipal HAZMAT Firefighters (n=340) | Expert consensus workshop criteria | 33.33% (7.49% – 70.07%) | 94.86% (91.90% – 97.00%) | Agreed with 3 of 9 physician unfit determinations; criteria missed 6 of 9 unfit cases. |
1. The Risk of Single-Cohort Thresholds
In the oil-and-gas cohort, 19 of 24 unfit examinations were identified. However, because the cut-off value was optimized on that specific dataset without separate validation on a new worker population, statistical over-fitting remains likely. In safety-critical operational environments, missing 5 of 24 unfit workers warrants caution.
2. High Specificity with Low Sensitivity
In the firefighter cohorts, criteria demonstrated high specificity (>94%) but low sensitivity (11% to 33%). The criteria cleared individuals whom experienced occupational physicians evaluated as unfit for heavy protective gear and toxic environment operations.
Contextual Demands vs. Static Thresholds
Automated scoring rubrics face limitations because occupational fitness depends on specific workplace exposures and physical demands, rather than isolated laboratory values alone.
For example, an elevated blood pressure reading or impaired fasting glucose carries distinct operational implications depending on the role:
- For an administrative office employee, these findings warrant medical referral, lifestyle modification, and scheduled monitoring, but rarely immediate job restriction.
- For a commercial crane operator or offshore worker, identical physiological readings combined with sudden dizziness risks or sedative medications present direct operational safety considerations.
Static scoring models evaluate fixed numeric thresholds. They cannot observe job ergonomics, functional task performance, or individual clinical context.
Micromeet builds governed healthcare AI infrastructure for continuous care. Clinicians and authorized staff review consequential outputs.
Regulatory and Professional Standards in Indonesia
Medical assessment standards in Indonesia establish clear governance boundaries:
- Medical Authority under Law No. 17 of 2023 on Health: Official health certificates and occupational fitness determinations are legal medical acts reserved for certified examining medical professionals holding valid registration and licensing under statutory health provisions verified at JDIH BPK. Software tools cannot independently issue legal medical determinations.
- Clinical AI Governance: Under health innovation frameworks evaluated by the Ministry of Health at JDIH Kemkes, technology serves as an administrative and analytical aid. Systems must maintain clear audit trails showing that drafts are reviewed and confirmed by licensed physicians.
- Personal Data Protection (Law No. 27 of 2022): Employee health records constitute specific personal data. Utilizing third-party software requires secure infrastructure, role-based access, and verifiable clinical data governance.
Safe Technology Roles: Administrative Drafting, Clinical Decision
The evidence indicates that technology should support, rather than replace, clinical review.
Examining doctors spend substantial administrative time compiling multi-source lab data, matching units, identifying typos, and drafting standard summaries.
A governed workflow divides tasks clearly:
- Administrative AI Support: Compiles multi-format test sheets, normalizes measurement units, flags out-of-range parameters, and prepares draft summaries with direct links to source data.
- Physician Judgment: Assesses operational job risks, evaluates clinical context, decides fitness categories, and signs official documentation.
Through MCU CoPilot in the Micromeet Care Workspace, technology organizes the documentation, while clinical determinations remain entirely with the physician. AI writes. Doctors decide.
Frequently Asked Questions (FAQ)
Can software legally issue an "unfit" determination for an employee in Indonesia?
No. Under Law No. 17 of 2023 on Health, health assessments and fitness certificates must be evaluated and authorized by certified medical practitioners. Software functions solely as an administrative drafting aid.
What is the primary safety consideration when evaluating automated scoring tools?
The primary clinical risk is low sensitivity (false negatives), where automated formulas clear workers who have subtle medical risks that an examining physician would identify through functional and clinical evaluation.
How can occupational clinics adopt AI tools compliantly?
Clinics can utilize AI for data normalization, anomaly flagging, and structured drafting, ensuring an audit log records physician review before reports are published.
