Installation

Set up the environment with Python 3.12:

conda create -n xxx python=3.12

Install additional dependencies:

pip install -r rq.txt

Verify the intalled environment

Here is the 5-fold cross validation results on AREDS Dataset with reading-center-defined phenotypes

Backbone	Fold	Phenotypes	Demographic	Age (Loss)	C‑Index
VMamba	fold‑1	Yes	smk	Pred (MAE + CoxPH)	0.878
VMamba	fold‑2	Yes	smk	Pred (MAE + CoxPH)	0.869
VMamba	fold‑3	Yes	smk	Pred (MAE + CoxPH)	0.883
VMamba	fold‑4	Yes	smk	Pred (MAE + CoxPH)	0.886
VMamba	fold‑5	Yes	smk	Pred (MAE + CoxPH)	0.882
VMamba	all	Yes	smk	Pred (MAE + CoxPH)	0.888

Start

Data Preparation

To train and evaluate the AREDS test set with the Cox model
(d_age_cox.py, cox function — see line 423), the input table must contain the following columns:

Category	Field(s)	Notes / Expected format
Demographics	age, smk	- age: numerical (chronological age) - smk: smoking status
Reading-center-defined phenotypes	GEOAWI, DRARWI, RPEDWI, INCPWI	Four Reading-center-defined phenotypes required by the model
Specialist-defined phenotypes	DRUSENQ, PIGDRUQ	Two Specialist-defined phenotypes required by the model

Ensure all required fields are present and free of missing values; otherwise
the Cox routine will raise a validation error.

FundusAge Model

To generate biological age values:

1. Prepare Checkpoints

   • The five-fold trained models can be accessed via the following Google Drive link：
   • Place all model checkpoints in the ./result_p directory.

2. Run Prediction
   Execute the following command to get predictions for each fold (X = 1–5):

   python d_age_cox.py --train_fold X

   After successful execution, a file named fold_X_test.csv will be generated in the root directory.

3. Data Matching by ID The image loading is based on patient/image IDs. Please ensure you prepare the correct CSV file with IDs matching the image filenames, I get the info using get_id() function in data_id_age_cox.py

   Modify the Following Lines If Needed. To ensure correct image loading paths, you may need to adjust:

   • d_age_cox.py: lines 562, 566

   • dataset_age_cox.py: line 33

   For example, image loading is defined as:

   X = Image.open('/prj0129/mil4012/AREDS/AMD_224/' + ID[:-4] + '.jpg')

   Make sure this path and file format match your local data setup.

4. Run Cox Evaluation on Predicted Results

After generating all 5 folds of test results, run cox_test.py to evaluate each fold's Cox performance.

• Use the test_fold variable at line 280 of cox_test.py to specify which fold to evaluate:
  • test_fold = 1 to 5: evaluate individual folds
  • test_fold = 0: evaluate the ensemble result by combining predictions from all five folds
  • see our biological age results and data spilt for AREDS in ./result_v3
  • Update: The biological age results and data spilt for AREDS2 are available in ./Areds2_results.

# Example (cox_test.py, line 280)
test_fold = 0  # use 0 for ensemble prediction

• Building and test with Reading-center based Cox model:

  python cox_test.py

• Building and test with Specialist based Cox model:

  python cox_doc.py

• Replace biological age with chronological age:
  • set bio = False in cox_test.py/ cox_doc.py line 305