| --- |
| license: mit |
| tags: |
| - regression |
| - medical-imaging |
| - fingernail |
| - hemoglobin |
| --- |
| |
| # pockethb-base |
|
|
| base hemoglobin regressor from the [pocketHb](https://github.com/jayanthvee/pocketHb) project. **not a measurement tool. read this whole card before using.** |
|
|
| ## what this is |
|
|
| a global, population-level Hb regressor on fingernail photos. methodology follows Rudokaite et al., BNAIC 2025 (Tilburg/Sanquin smartphone-Hb paper): |
|
|
| - frozen convnext_tiny.fb_in22k_ft_in1k via imm (no fine-tuning) |
| - Shades-of-Gray (p=6) illumination correction → 224x224 resize → ImageNet normalize |
| - per-patient aggregation: mean + std of crop embeddings (1536-d patient vector) |
| - standardize → PLS + SVR(RBF) → isotonic calibration → weighted blend |
| - trained on Nature Sci Data 2024 fingernail+Hb dataset (n=250 subjects, single CBC each) |
|
|
| trained on **nail crops only**. the public Nature 2024 release ships 600x800 images, but 606 of 750 skin bboxes were labelled in a taller source frame and now sit below the bottom edge of the released images. only 83 patients would survive a nail+skin fusion — we ran nail-only on the full 250 instead. |
|
|
| ## metrics (5-fold stratified CV, n=250) |
|
|
| - **MAE: 2.09 g/dL** |
| - **RMSE: 2.74 g/dL** |
| - **R²: -0.05** |
|
|
| R² ≈ 0 means this global model is essentially predicting near the dataset mean. it barely beats the predict-mean floor (MAE ~2.14 g/dL) on this dataset. **this is the documented ceiling of population-level Hb modeling at sub-1000-subject scale**, not a bug: |
|
|
| - The Rudokaite/BNAIC paper itself, with the same methodology on n=159, got MAE 0.6 mmol/L — that sounds clinical until you notice their predict-mean floor was ~0.63 mmol/L. similar R² ≈ 0 story, smaller absolute number because their donor population had Hb std 0.79 mmol/L vs our 2.67 g/dL. |
| - Mannino et al. (PNAS 2025) only got clinical-near accuracy with n=9,061 subjects PLUS per-user personalization. their global model alone is not clinically useful either; they patented and locked the personalization layer (US 12268498). |
|
|
| ## intended use |
|
|
| **only as an input to a per-user calibration / personalization layer.** the open-source contribution of [pocketHb](https://github.com/jayanthvee/pocketHb) is exactly that personalization layer. without it, this base regressor is research artifact only. |
|
|
| ## not a medical device |
|
|
| research replication only. **must not be used to estimate anyone\'s actual hemoglobin in any clinical, diagnostic, or treatment context.** not FDA cleared. not validated clinically. not a doctor. if you need a hemoglobin reading, get a blood test. |
|
|
| ## how to use |
|
|
| `python |
| import pickle |
| from huggingface_hub import hf_hub_download |
| |
| # download bundle |
| bundle_path = hf_hub_download(repo_id='bubbaonbubba/pockethb-base', filename='pockethb_base.pkl') |
| with open(bundle_path, 'rb') as f: |
| bundle = pickle.load(f) |
| |
| # bundle contains: backbone_name, image_size, imagenet_mean/std, |
| # shades_of_gray_p, modalities, aggregation, blender, n_train_patients, seed |
| ` |
| |
| see the [pocketHb repo](https://github.com/jayanthvee/pocketHb) for the full inference pipeline including image preprocessing and embedding extraction. |
| |
| ## citation context |
| |
| - Mannino et al., PNAS 2025. doi:10.1073/pnas.2424677122 |
| - Rudokaite et al., BNAIC 2025 (Tilburg/Sanquin) |
| - Nature Sci Data 2024 dataset: doi:10.1038/s41597-024-03895-9 |
| |
| ## license |
| |
| MIT. |
| |
