Research and Development
SMAILE develops and maintains cutting-edge open-source tools that advance medical AI research and clinical practice.
Research Tools and Platforms
SMAILE develops and maintains innovative research tools that advance the field of medical AI while remaining freely accessible to the global research community. Our platforms embody our commitment to open science and collaborative innovation, providing researchers and clinicians with sophisticated tools for medical AI research, clinical pathway analysis, and medical imaging applications.
MEDLEY (Medical Ensemble Diagnostic system with Leveraged diversitY) represents a paradigm shift in medical AI. Rather than seeking a single 'perfect' model, MEDLEY orchestrates multiple AI systems in parallel, preserving their diverse perspectives, biases, and disagreements as valuable clinical insights.
In a large language model demonstrator, unlike traditional approaches that collapse multiple AI outputs into a single answer, MEDLEY treats disagreement as an informative signal and bias as a form of specialization. Our proof-of-concept demonstrates how 30+ language models can collaborate to surface diagnostic uncertainty, identify rare conditions, and provide clinicians with a range of perspectives rather than a single recommendation.
Key Benefits
- Preserves minority perspectives that might catch rare diseases
- Makes AI bias transparent and manageable
- Reduces automation bias by encouraging active clinical reasoning
- Mirrors multidisciplinary tumor boards in practice
Purpose and Innovation
Multi-model medical AI ensemble system that orchestrates 30+ diverse AI models. Revolutionary approach that preserves disagreement as diagnostic insight rather than collapsing into consensus. Treats bias as specialization and makes uncertainty visible to clinicians.
Key Capabilities
- Parallel orchestration of heterogeneous models.
- Bias detection through model disagreement patterns.
- Uncertainty quantification and calibration.
- Transparent provenance documentation.
- Population-specific disease recognition.
Learn More
HealthProcessAI is a comprehensive, dual-language framework (Python & R) that applies process mining techniques to healthcare data, incorporating integrated AI capabilities. Developed at SMAILE, it helps researchers and clinicians understand actual care pathways, identify bottlenecks, and optimize clinical workflows.
Unique Features
- Parallel Python and R implementations for flexibility
- LLM integration for AI-powered clinical insights
- Multi-model orchestration consolidating insights from Claude, GPT-4, Gemini, and more
- Comprehensive tutorials for healthcare professionals
Purpose and Innovation
Dual-language (Python & R) framework for healthcare process mining with integrated AI-powered insights. Unique orchestrator feature consolidates insights from multiple LLM models into unified comprehensive reports for clinical pathway optimization
Key Capabilities
- Complete 5-step analysis pipeline
- LLM integration (Claude, GPT-4, Gemini, DeepSeek, Grok)
- Multi-model insight orchestration
- Sepsis tracking & pathway optimization
- Clinical-friendly tutorials
Access
HealthProcessAI - Process Mining Framework | SMAILE, Karolinska Institutet
SpekPy is a state-of-the-art toolkit for estimating photon spectra from X-ray tubes. Originally available only as a Python package requiring programming expertise, SpekPy Web now provides both a user-friendly graphical interface and an API, making advanced X-ray modeling accessible to all researchers and medical physicists.
Technical excellence
- Database of 445 materials (all elements H-U plus phantom materials and tissues).
- Validated against International Bureau of Weights and Measures standards.
- All estimates within 3.5 percent agreement.
- Average calculation time: 2.5 seconds per spectrum.
Perfect for
- Medical physicists performing dosimetry.
- Researchers studying radiation effects.
- Equipment developers optimizing x-ray systems.
- Anyone needing accurate spectrum estimation without coding.
Purpose and Innovation
Advanced x-ray spectrum estimation toolkit with both GUI and API access. Makes sophisticated photon spectrum modeling accessible without programming expertise. Validated against international standards with exceptional accuracy.
Key Capabilities
- 445 material database (H-U elements, tissues).
- Web-based GUI for easy access.
- RESTful API for programmatic use.
- 3.5 percent accuracy vs. BIPM standards.
- 2.5 second average computation.
trustcv (Trustworthy Cross-Validation) addresses a pervasive methodological problem in medical machine learning: improper model validation. Recent systematic reviews have documented widespread methodological flaws in ML publications, with data leakage identified in hundreds of studies across multiple scientific fields. These errors frequently arise from violations of the independence assumption inherent in standard cross-validation, producing inflated performance estimates that fail to generalise to clinical deployment. Rather than relying on generic splitting strategies that ignore the structure of clinical data, trustcv provides 29 specialised cross-validation methods that respect the hierarchical, temporal, and spatial characteristics of medical datasets.
Medical data routinely violates the independent and identically distributed (i.i.d.) assumption through patient-level dependencies (i.e., multiple measurements from the same individual), longitudinal correlations (e.g., ICU monitoring, disease progression), and geographic clustering (e.g., epidemiological surveillance). Standard k-fold cross-validation applied naively to such data can dramatically overestimate model performance; for example, studies have shown apparent accuracy dropping from 95% to 60% once proper patient-level splitting is applied. trustcv makes these failure modes explicit and preventable through automatic leakage detection and method-appropriate data splitting.
Key Benefits
- Prevents six types of data leakage common in medical ML (patient, temporal, spatial, preprocessing, duplicate, and feature-target leakage)
- Framework-agnostic design supporting scikit-learn, PyTorch, TensorFlow, MONAI, and JAX
- Reduces validation implementation effort while enforcing methodological rigour
- Provides regulatory compliance features aligned with FDA and CE MDR requirements for AI-based medical devices
Purpose and Innovation
A comprehensive, open-source Python toolkit that implements 29 cross-validation methods organised into four categories based on the data characteristics they address: i.i.d. methods (n=9), grouped/hierarchical methods (n=8), temporal methods (n=8), and spatial methods (n=4). The toolkit combines a systematic review of validation methodology with a reference implementation, serving both as a practical tool and as an educational resource for proper validation practice.
Key Capabilities
- 29 specialised cross-validation methods covering all major medical data structures.
- Automatic detection of six distinct data leakage types before model evaluation.
- Clinical metrics with confidence intervals (e.g., sensitivity, specificity, PPV, NPV).
- Regulatory documentation templates mapped to FDA GMLP and CE MDR frameworks.
- Interactive tutorials and worked examples across multiple clinical domains.