Started a new project- creating a Mamba Hybrid Prototype

Models
1

I started a new project and would like to share with the community- please keep in mind I am still in my early stage of progress and fails. If anyone is interested, here is my readme. Happy to hear your thoughts. all the best Katharina

Hamburg SafetyMamba Hybrid Prototype

This project explores a human-centered hybrid model architecture for safety prediction in
emergency response scenarios. Inspired by—but not replicating—the goals of large-scale systems
like APONA, this prototype takes a more compassionate, context-sensitive approach to patient and
staff care.

Overview

We combine modern sequence modeling (GRU/Mamba-based backbones) with structured
multi-source tabular data from clinical and operational domains. The goal is to move beyond
logistical optimization and ensure crisis detection, staff burnout prediction, delay estimation, and
safety outcomes are centered in training.

Goals

  • Model patient-centered safety outcomes in Hamburg’s EMS setting
  • Predict multivariate targets including:
  • Crisis level
  • Burnout risk
  • Retention probability
  • Response delay
  • Handoff quality
  • Adverse event risk
  • Evaluate hybrid backbones with modern encoders (e.g., Mamba/GRU)
  • Maintain full transparency in the modeling process, including failures and fixes

Project Motivation

This project is born from the belief that:

“Patients and frontline staff should not be reduced to parameters in a logistic equation.”
It is designed with care, not just for performance.

Data

  • Synthetic and real-world structured time-series data
  • Patient vitals, demographics, clinical conditions
  • Crew stress, fatigue, and shift conditions
  • Environmental and systemic stressors

Model

A hybrid safety model with:

  • Modular encoders for clinical, operational, temporal, and environmental data
  • Optional backbone: GRU (currently active) or Mamba (planned)
  • Multi-task output heads with learned task uncertainty

Development Stages

  • ■ Prototype v1: Fully working model with synthetic data
  • ■ Hybrid encoder built and tested
  • ■ Real Hamburg-style data wired and flowing through full pipeline
  • ■ Forward pass verified
  • ■ Training loop executing on real data
  • ■ Next: Model evaluation and interpretability

Links

1 Like
2

Shift in Scope

After two days of exploratory work, the direction of the project was clarified: shifted from an full-behavioral simulation to a **simplified, hard-variable-only coordination pipeline**.

This means:

- No speculative modeling of staff psychology, inter-departmental friction, or refusal behavior (yet) -

Only **observable, processable variables** (e.g., triage level, ORBIS/IVENA capacities,

timestamps) - Focus on **routing, timing, and discharge coordination**

This change enables:

- **Faster iteration and clearer debugging**

  • More **transparent results** and easier community validation
  • A stable foundation for layering in behavioral complexity later

## ■ Why Mamba Was Not Used

The project initially considered newer structured state space models (SSMs) like **Mamba** for

their ability to model long-range dependencies. After review, the decision was made to **not pursue Mamba** at this stage:
■ Sequence length → Patient journeys are short, bounded sequences (EMS → ED → Ward → Discharge)
■ Interpretability -GRUs are easier to debug and explain, especially in healthcare contexts
■ Complexity -Mamba adds hyperparameter tuning and potential instability, with no meaningful gain in this setup

The model will instead use a **hybrid GRU + MLP** structure, which is appropriate for the sequence length, modeling requirements, and training simplicity.

## ■ Current Focus

building a **Minimal Viable AI Assistant** that:

  • Simulates patient pathways using synthetic, realistic hospital data
  • Assists in coordination decisions (e.g., hospital routing, ward assignment, discharge comms)
  • Evaluates outcomes like:

**Pathway latency**

  • **Ward suitability**
  • **Discharge coordination completeness**

All assumptions are stated clearly and transparently, and all data is simulated (no real patient data).

## ■ Contributions Welcome

If you’re working on hospital flow, medical AI ethics, synthetic healthcare datasets, or practical

coordination tools — contributions, critique, and validation ideas are all welcome.

This project will remain open-source, community-aligned, and accountable to clinical realism.

1 Like