File size: 3,353 Bytes
31be935 ac536aa 31be935 ac536aa 31be935 ac536aa 31be935 1739a04 31be935 1739a04 31be935 1739a04 31be935 ac536aa 31be935 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | ---
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.
|