QinEmPeRoR93

Agentic-TimesFM-AKI

README

License: apache-2.0

📌 Model Overview

Agentic-TimesFM-AKI is a specialized, privacy-preserving multi-modal clinical framework designed for the continuous prediction of synergistic nephrotoxicity — specifically Acute Kidney Injury (AKI) induced by the concurrent administration of Vancomycin and Piperacillin-Tazobactam (Zosyn).

This repository hosts pre-trained weights, adapters, and quantized GGUF variants for the dual-model system, integrating:

Gemma-4 12B Sentinel (QLoRA fine-tuned on synthetic clinical narratives).
TimesFM 2.5 Agent (LoRA fine-tuned zero-shot time-series forecaster).

🏗️ Architecture & Functionality

The system leverages a dual-agent orchestration framework:

TimesFM 2.5 Forecast Engine: Ingests longitudinal lab values (e.g., serum creatinine, BUN) and projects future 72-hour creatinine trajectories.
Gemma-4 12B Clinical Sentinel: Receives TimesFM's 72-hour projections alongside patient demographics and medication timelines to output a structured binary prediction (AKI_POSITIVE / AKI_NEGATIVE) and a natural language clinical warning.

text
Patient EHR / Labs ──▶ TimesFM 2.5 Agent (72h Forecast)
                              │
                              ▼
Patient Context ──────▶ Gemma-4 12B Sentinel ──▶ Binary Risk + Clinical Warning

📊 Evaluation & Metrics

The framework was trained exclusively on differentially private synthetic data ( $ε = 10$ ) to preserve patient privacy and validated on a real-world eICU holdout cohort (N=200):

Table with columns: Metric, Point Estimate, 95% Bootstrap Confidence Interval
Metric	Point Estimate	95% Bootstrap Confidence Interval
Accuracy	0.970	0.945 – 0.990
Sensitivity (Recall)	0.944	0.892 – 0.988
Specificity	0.991	0.971 – 1.000
Precision	0.988	0.961 – 1.000
F1-Score	0.966	0.934 – 0.989

Performance Comparison vs. Baselines (eICU Holdout)

Agentic-TimesFM-AKI (F1: 0.966) significantly outperformed traditional baselines:
- Random Forest (F1: 0.752, $p < 0.001$ )
- XGBoost (F1: 0.748, $p < 0.001$ )
- Logistic Regression (F1: 0.700, $p < 0.001$ )

⚠️ Important Considerations & Limitations

Formatting Fragility (Domain Shift): External validation on the MIMIC-IV demo cohort revealed performance degradation ( $F 1 = 0.421$ ) due to structural schema shifts. Prompt recalibration and strict schema harmonization are required before multi-center deployment.
Generation Token Allocations: Truncating generation tokens (e.g., < 100 tokens) prematurely cuts off the clinical reasoning chain, causing false negatives. Dynamic or high token limits ( $\geq 250$ tokens) are recommended.
Clinical Scope: Currently optimized specifically for Vancomycin + Piperacillin-Tazobactam synergistic risk assessment.

💻 Quick Start & Usage

python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "QinEmPeRoR93/Agentic-TimesFM-AKI"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

prompt = """<|turn>user
Patient: 68yo male, MAP: 75 mmHg, Vancomycin Trough: 18 mg/L, Piperacillin-Tazobactam: Active.
TimesFM 72h Serum Creatinine Projection: [1.2, 1.4, 1.9 mg/dL]
Assess synergistic nephrotoxicity risk.
<|turn>model
"""

inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=250)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

⚖️ Licensing & Terms

Code & Model Weights: Released under the Apache 2.0 License.
Base Model License: Fine-tuned from Google's Gemma 4. Usage must conform with Google Gemma Terms of Use.

📄 Citation

If you use this model or code in your work, please cite the corresponding paper:

bibtex
@article{alsakkaf2026agentictimesfmaki,
  title   = {Agentic-TimesFM-AKI: A Dual LLM–Time Series Framework for Predicting Drug-Induced Acute Kidney Injury with Privacy-Preserving Synthetic Data},
  author  = {Alsakkaf, Gamal Esam Ahmed},
  journal = {medRxiv},
  year    = {2026},
  doi     = {10.64898/2026.07.30.26359271},
  url     = {https://www.medrxiv.org/content/10.64898/2026.07.30.26359271v1}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

QinEmPeRoR93

Model Tree

Base

this model

Input Modalities

TextAudioImageVideo

Output Modalities

Text

Supported Functionality