FairHealth: An Open-Source Python Library for Trustworthy Healthcare AI in Low-Resource Settings

The paper introduces FairHealth, an open-source Python library designed to bridge critical gaps in healthcare AI for low-resource settings by providing a unified, modular framework that integrates fairness auditing, privacy-preserving federated learning, low-bandwidth explainability, and specialized tools for Global South datasets.

The Gist

Imagine you are building a team of doctors to help people in remote villages where internet is spotty, electricity is unreliable, and there aren't many specialists. You want to use a computer program (AI) to help them, but you have three big worries:

1. Is it fair? Does the computer treat a young woman from a village the same way it treats an older man from the city?
2. Is it safe? Can we teach the computer without stealing private patient records?
3. Can we trust it? If the computer makes a suggestion, can a local nurse understand why it made that choice without needing a PhD in math?

FairHealth is a new, free "toolbox" (a Python library) designed specifically to solve these three problems for places like Bangladesh and other low-resource countries. Think of it as a Swiss Army Knife for ethical healthcare AI.

Here is how the toolbox works, broken down into its six main tools:

1. The "Fairness Mirror" (fairhealth.fairness)

The Problem: Often, AI models are trained on data from wealthy countries. When you use them in a different place, they might get it wrong for certain groups of people (like women or specific ethnic groups). It's like a weather app trained only on London weather trying to predict rain in the Sahara; it just won't work.
The Tool: This module acts like a mirror that checks if your AI is being biased. It runs a "fairness audit" to see if the AI treats different groups equally.

• Real-world example: The paper shows that without this tool, an AI checking heartbeats (ECG) was only fair 23% of the time between men and women. After using this tool to "fix" the AI, fairness jumped to 71%.

2. The "Translator" (fairhealth.explain)

The Problem: Most AI is a "black box." It gives an answer, but no one knows how it got there. In a busy clinic in a low-resource setting, a doctor can't ask a computer scientist to explain the math. They need a simple reason.
The Tool: This module translates complex math into plain language rules, like a translator speaking to a local elder.

• Real-world example: Instead of saying "The probability score is 0.88," it says, "Rule 1: High Blood Pressure AND High Blood Sugar = High Risk." A study mentioned in the paper found that doctors preferred these simple "rule-based" explanations over complex charts.

3. The "Secret Vault" (fairhealth.federated)

The Problem: Hospitals can't share patient records because of privacy laws. It's like trying to teach a chef a new recipe by sending them the actual ingredients, but the ingredients are locked in a vault.
The Tool: This tool uses a special kind of "magic lock" (called Homomorphic Encryption). It allows hospitals to train the AI together without ever opening the vault or sending the actual patient data. They only send "encrypted hints" about the recipe.

• The Result: The paper claims this method shrinks the amount of data sent over the internet by 97.5% (making it fast even on slow connections) while keeping the data mathematically unbreakable by hackers.

4. The "Emergency Triage" (fairhealth.lowresource)

The Problem: During disease outbreaks (like Dengue fever), clinics get overwhelmed. They need a quick way to sort patients, but the system must work offline and speak the local language.
The Tool: This is a smart sorting assistant for Dengue fever. It asks simple questions (Age, Location, Housing type) and gives a recommendation in English or Bengali.

• Real-world example: If a child in Dhaka has a fever, the tool can instantly say, "Severe: Go to the doctor immediately," helping doctors decide who needs help first.

5. The "Equity Compass" (fairhealth.equity)

The Problem: When disasters happen (like floods), aid often goes to the places that are easiest to reach (cities), leaving the hardest-hit rural areas behind. Old AI models just copy this mistake.
The Tool: This module acts as a compass that points toward the people who need help the most, regardless of where they live. It uses a special technique to ignore "location bias."

• Real-world example: In the 2022 Bangladesh floods, this tool changed the priority list. A rural area called Sunamganj, which was previously ranked 14th for aid, was correctly moved up to Rank 1 because the model realized they were suffering the most.

6. The "Open Library" (fairhealth.datasets)

The Problem: Most medical AI research requires special permission (a "Data Use Agreement") to access patient records. This locks out independent researchers, students, or people in countries without big hospital networks.
The Tool: FairHealth is the first toolbox that only uses data that is already free and public. You don't need to ask for permission or sign legal papers.

• The Benefit: Anyone with a computer can download the data and start building fair AI immediately.

Summary

FairHealth is a free, open-source toolkit that helps researchers and doctors build AI that is fair (doesn't discriminate), private (keeps secrets safe), and explainable (easy to understand). It is built specifically for the challenges of low-resource settings, using only data that is free for everyone to use.

You can install it just like any other app (pip install fairhealth) and start using these tools to make healthcare AI safer and more trustworthy for everyone.

Technical Summary

Technical Summary: FairHealth

Problem Statement

Machine learning holds significant promise for healthcare, yet its real-world deployment, particularly in low-resource and low-income country (LMIC) settings, is hindered by three structural limitations:

1. Demographic Bias: Models trained on population-level data often underperform for minority groups. For instance, uncorrected ECG-based myocardial infarction detection models exhibit a disparate impact ratio of 0.23 across sex groups, far below the 0.80 threshold for equity.
2. Privacy Constraints: Legal protections on healthcare records prevent data sharing across institutions. While federated learning (FL) offers a solution, existing Python libraries lack integration of cryptographic homomorphic encryption (HE) accessible to healthcare researchers.
3. Explainability Gaps: Clinical decision support tools in LMICs (e.g., Bangladesh) require explanations interpretable by clinicians without ML expertise, operation in low-bandwidth environments, and support for local languages. Existing libraries (e.g., SHAP, LIME) lack clinical workflow integration.

Current toolkits like PyHealth cover broad EHR datasets but omit fairness auditing, federated learning, and LMIC-specific deployment. Conversely, fairness libraries like AIF360 lack healthcare-specific datasets, and FL frameworks like FATE lack healthcare focus and impose high engineering overhead.

Methodology and Library Design

FairHealth is an open-source, modular Python library designed to address these gaps. It is built upon five peer-reviewed research contributions and structured into six loosely coupled modules. The library adheres to three core design principles:

• Public Data Only: All datasets are downloadable from public sources without institutional data use agreements (DUAs). Crucially, the regional healthcare data used in the maternal health module—specifically the Healthcare_Access_Score and MMR (Maternal Mortality Ratio per 100k) variables—are sourced from real-world public registries, not synthetic or simulated values. These indicators are explicitly drawn from two authoritative sources: (1) the DGHS Dashboard 2024 (Directorate General of Health Services, Bangladesh) and (2) the UNFPA MPDSR 2023 (Maternal and Perinatal Death Surveillance and Response).
• Paper-Anchored: Every module is tied to a specific peer-reviewed publication, with key results documented in docstrings for traceability.
• Clinical Framing: Outputs are presented in clinical language rather than ML jargon, validated by a study involving 14 healthcare professionals.

Core Modules

1. fairhealth.fairness (Biosignal Fairness):

   • Function: Provides model-agnostic fairness metrics (Demographic Parity, Equalized Odds, Disparate Impact) for biosignals and tabular data.
   • Innovation: Includes intersectional_fairness to evaluate multiple sensitive attributes simultaneously (e.g., sex × age).
   • Validation: Demonstrated on the PTB-XL dataset, where adversarial debiasing improved the disparate impact ratio from 0.23 to 0.71 while maintaining AUROC at 0.8472.

2. fairhealth.explain (Hybrid Explainability):

   • Function: Combines ante-hoc fuzzy logic rules with post-hoc SHAP explanations.
   • Innovation: Addresses the "trust deficit" in resource-constrained settings. A clinician validation study showed 71.4% preference for this hybrid approach over SHAP-only or score-only explanations.
   • Performance: Achieved 88.67% accuracy on the UCI Maternal Health Risk dataset, outperforming Gradient Boosting baselines by 2.46 percentage points.

3. fairhealth.federated (Privacy-Preserving FL):

   • Function: Implements federated learning with homomorphic encryption (HE) and differential privacy (DP).
   • Innovation: Uses the MedHE approach, co-designing adaptive gradient sparsification with CKKS homomorphic encryption.
   • Performance: Reduces communication overhead by 97.5% (from 1,277 MB to 32 MB) while maintaining macro-F1 scores statistically equivalent to standard FedAvg. It improves resistance to membership inference attacks (MIA) to near-random levels (51.1% success rate).

4. fairhealth.lowresource (Multilingual Triage):

   • Function: A Decision Tree classifier for dengue triage supporting English and Bengali.
   • Innovation: Operates in low-bandwidth conditions using demographic features (Age, Gender, AreaType, etc.).
   • Performance: Achieved an F1 score of 0.802 and AUC of 0.851. A pilot study (n=50) reported 75% user satisfaction with its confidence threshold mechanism (rerouting low-confidence cases to doctors).

5. fairhealth.equity (Disaster Aid Allocation):

   • Function: An adversarial debiasing architecture for equitable flood aid distribution.
   • Innovation: Uses gradient reversal layers to learn district-invariant vulnerability representations, correcting urban biases in historical data.
   • Performance: Evaluated on 87 upazilas in Bangladesh, the model reduced the statistical parity difference by 41.6% and regional fairness gap by 43.2% with a minimal R² cost (2.7 percentage points).

6. fairhealth.datasets (Public Loaders):

   • Function: Runtime loaders for six specific datasets, including Bangladesh-specific maternal health (sourced from DGHS Dashboard 2024 and UNFPA MPDSR 2023), dengue surveillance, and flood damage (PDNA) data, alongside standard datasets like PTB-XL and MIMIC-III Demo.

Key Results and Contributions

The paper claims the following contributions:

1. Unified Framework: The first pip-installable library integrating federated learning, fairness, explainability, and low-resource tools specifically for healthcare.
2. Reproducibility: The first healthcare AI toolkit built entirely on publicly available datasets, removing the barrier of institutional DUAs for independent researchers.
3. LMIC Focus: A curated collection of Bangladesh-specific health datasets (maternal, dengue, flood) not found in existing ML libraries, featuring real-world indicators like MMR and Healthcare_Access_Score.
4. Open Implementation: Open-source code for five peer-reviewed methods, enabling direct reproduction and extension.

Technical Performance Highlights:

• Fairness: Improved disparate impact ratio from 0.23 to 0.71 on ECG data.
• Privacy: Achieved 97.5% communication reduction in FL with HE while maintaining model utility (macro-F1 ≈ 0.95).
• Explainability: 71.4% clinician preference for hybrid Fuzzy+SHAP explanations.
• Equity: Shifted priority rankings for 70.6% of disaster-affected regions, significantly up-ranking vulnerable rural areas (e.g., Sunamganj).

Significance and Claims

The paper positions FairHealth as a complementary layer to existing broad-coverage libraries like PyHealth, specifically targeting the "trustworthiness dimension" (fairness, privacy, explainability) that other tools leave open. Its primary significance lies in democratizing healthcare AI research for the Global South by:

• Eliminating the need for institutional data access agreements by utilizing verified public registries.
• Providing tools tailored for low-bandwidth and multilingual clinical environments.
• Ensuring that fairness and privacy are not afterthoughts but integrated, auditable components of the ML workflow.

The authors state that the library enables researchers worldwide, including those without hospital affiliations, to conduct rigorous, reproducible healthcare AI research. Future work is limited to expanding the federated module with full TenSEAL-based CKKS encryption for neural network weights and integrating real-time dashboards for dengue surveillance.