Software @software0408

ML/AI Engineer & PI. Building a Nigerian chest X-ray benchmark to measure how well Western AI models work on African patients. linkedin.com/in/software040… Lagos, Nigeria Joined March 2019

Tweets

49
Followers

27
Following

216
Likes

541

Software @software0408

2 months ago

x.com/i/article/2043…

0 0 1 4 0

View Details

The lesson: don't choose rules OR transformers. Use both. Rules give you interpretable precision. BERT gives you semantic reach. Our ensemble: rules primary on 11 labels, BERT primary on 2. 495 reports ready for radiologist validation. #MedicalAI #NLP #ChestXray #Africa #BuildInPublic

0 0 1 12 0

View Details

Software @software0408

2 months ago

Also verified negation detection across 5,517 reports. 0 failures out of 134 negated mentions. And reviewed 617 reports with 3+ labels — 89.9% of all label assignments supported by text evidence.

1 0 1 8 0

View Details

Software @software0408

2 months ago

Week 10/48. This week I audited our entire NLP pipeline. The question: should you use rules or transformers for medical text? The answer surprised me. Thread.

1 0 1 12 0

View Details

Software @software0408

2 months ago

@heyrimsha This is great

0 0 0 101 0

View Details

MIT CSAIL @MIT_CSAIL

2 months ago

A prompt you can use in Claude to help build & launch apps, v/@milesdeutscher.

16 291 2K 203K 2K

View Details

Software @software0408

2 months ago

@ellen_in_sf literature review

0 0 0 14 0

View Details

Software @software0408

2 months ago

The pipeline: two extractors working in parallel. Rule-based: 447-phrase bilingual dictionary with negation detection BioClinicalBERT: fine-tuned transformer, Normal as 14th label Rules handle 11 labels. BERT handles 2 that rules can't detect. They vote. Flag disagreements.

0 0 0 8 0

View Details

Software @software0408

2 months ago

@lilyyang169 community

0 0 0 7 0

View Details

Software @software0408

2 months ago

The dataset: 5,517 chest X-rays from FMC Ebute-Metta, Lagos. Free-text radiology reports. English mixed with local medical terms. 48% had no section headers. Step 1: turn those reports into structured labels 13 pathologies + Normal. That took Weeks 1–8.