Agentic-TimesFM-AKI

A Dual LLM–Time Series Framework for Predicting Drug-Induced Acute Kidney Injury with Privacy-Preserving Synthetic Data

---

📌 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:

1. Gemma-4 12B Sentinel (QLoRA fine-tuned on synthetic clinical narratives).
2. 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.

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 ($\varepsilon = 10$) to preserve patient privacy and validated on a real-world eICU holdout cohort (N=200):

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

1. Formatting Fragility (Domain Shift): External validation on the MIMIC-IV demo cohort revealed performance degradation ($F1 = 0.421$) due to structural schema shifts. Prompt recalibration and strict schema harmonization are required before multi-center deployment.
2. 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 ($\ge 250$ tokens) are recommended.
3. Clinical Scope: Currently optimized specifically for Vancomycin + Piperacillin-Tazobactam synergistic risk assessment.

---

💻 Quick Start & Usage

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:

@article{saka2026agentictimesfmaki,
  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 E.},
  year    = {2026},
  note    = {Manuscript under review}
}