Applications of Machine Learning in Noninvasive Anemia Diagnosis: A Systematic Review Based on Cross-Industry...

Anemia is a disorder caused by insufficient red blood cell count or hemoglobin concentration, affecting nearly a quarter of the global population each year. An undiagnosed condition can escalate into life-threatening complications, particularly among pregnant women and children. Although anemia causes identifiable symptoms, accurate diagnosis typically requires invasive venipuncture with quality-controlled analyzers and reagents operated by trained personnel. This diagnostic procedure is laborious and costly in deprived regions and a less resource-demanding alternative would be beneficial. This systematic review aims at outlining the recent efforts for creating such noninvasive alternative procedures via machine learning (ML).

A search was conducted through PubMed, Web of Science, and Scopus via tailored queries, yielding 1923 entries. Following screening, 58 eligible records were included in the qualitative synthesis. The Cross-Industry Standard Process for Data Mining (CRISP-DM) phases guided the selection of the main characteristics of each paper.

The contributions span a variety of algorithms and input data. The most frequent acquisition sites and data types were eye conjunctiva images, fingertip PPG signals, and palmar area images, respectively. Estimation of hemoglobin was of more interest than anemia classification. The light source in most experiments was at low energy or near-infrared spectrum. The smartphone served as a convenient device for signal acquisition or model deployment via a mobile app.

Substantial progress has been achieved in ML applications for noninvasive anemia detection, leading to breakthroughs in prevention and diagnosis. Nonetheless, there were challenges to building robust procedures, such as data scarcity and imbalance, limited diversity, and confounding factors, some of which have been addressed recently.