AI-powered skin spectral imaging enables instant sepsis diagnosis and outcome prediction in critically ill patients

TL;DR

This study demonstrates that deep learning applied to a single, rapid hyperspectral skin imaging cube enables noninvasive sepsis diagnosis and 30-day mortality prediction in ICU patients. Palm-site HSI yielded AUROCs of 0.80 for sepsis and 0.72 for mortality, which improved substantially to 0.94 and 0.83, respectively, when incorporating structured clinical data collected within the first (and up to ten) hours of ICU admission. The approach outperformed widely used clinical biomarkers and scores, supporting its potential as a fast pre-screening tool, particularly in time-critical or resource-limited settings. External validation and multi-site studies are needed to confirm generalizability, but the method holds promise for rapid, mobile, noninvasive microcirculatory assessment to guide early intervention in sepsis care.

Abstract

With sepsis remaining a leading cause of mortality, early identification of patients with sepsis and those at high risk of death is a challenge of high socioeconomic importance. Given the potential of hyperspectral imaging (HSI) to monitor microcirculatory alterations, we propose a deep learning approach to automated sepsis diagnosis and mortality prediction using a single HSI cube acquired within seconds. In a prospective observational study, we collected HSI data from the palms and fingers of more than 480 intensive care unit patients. Neural networks applied to HSI measurements predicted sepsis and mortality with areas under the receiver operating characteristic curve (AUROCs) of 0.80 and 0.72, respectively. Performance improved substantially with additional clinical data, reaching AUROCs of 0.94 for sepsis and 0.83 for mortality. We conclude that deep learning-based HSI analysis enables rapid and noninvasive prediction of sepsis and mortality, with a potential clinical value for enhancing diagnosis and treatment.

Figures (11)

• Figure 1: We explore *hsi for automated, non-invasive and rapid sepsis diagnosis and mortality prediction. In a prospective study of over 480 *icu patients, we collected *hsi and *rgb images of the palm and annular finger, and clinical data. Deep learning accurately predicts sepsis and mortality from *hsi data, with improved performance when combined with clinical data. Our method outperforms widely used clinical biomarkers and scores such as the *qsofa score and the *sirs criteria.
• Figure 2: *hsi can rapidly and non-invasively diagnose sepsis and predict mortality.*roc are shown for sepsis diagnosis (A) and mortality prediction (B) models based on *hsi data (gold), stacked tissue parameter images (TPI, pink) and *rgb (violet) data of the palm (left) and annular finger (right).The shaded areas denote the 95% confidence interval across 1000 bootstrap samples, and mean and standard deviation of the *auroc are reported in the legend. Sample images of a septic (light red box) and non-septic (light green box) patient, as well as a survivor (dark green) and non-survivor (dark red) are included on the bottom right, with the circle denoting the annotated skin region.
• Figure 3: Septic patients and non-survivors possess significantly lower palm tissue oxygen saturation, and higher tissue haemoglobin and water index. The subfigures show the distribution of the functional parameters oxygen saturation, perfusion index, haemoglobin index and water index, derived from *hsi palm measurements, for septic and non-septic patients (A), and survivors and non-survivors (B). The boxes denote the quartiles of the distribution with the whiskers extending up to 1.5 times the interquartile range, and the median and mean drawn as solid and dashed lines, respectively. Each dot represents one patient. Tissue parameter index distributions for the measurement site finger are available in figure \ref{['fig:tpi_finger']}.
• Figure 4: Adding clinical data boosts the sepsis diagnosis and mortality prediction performance. The performance of sepsis diagnosis (A) and mortality prediction (B) using *hsi data of the palm (HSI palm model, gold), a combination of *hsi and clinical data (HSI palm + clinical data model, bronze), and clinical data alone (clinical data model, blue) is shown, categorised by data availability within one hour (left) and ten hours (right) from admission to the intensive care unit. Within the subplots, the performance of the HSI palm model is compared to HSI palm + clinical data and clinical data models that incorporate - from left to right - the most important, two most important, three most important or all clinical data features available within the specified timeframe of one hour or ten hours after *icu admission. The number of clinical data features used in the model is indicated in brackets. The ranking of the clinical data features according to feature importance was derived from the clinical data model through *rfe guyon2002gene starting from the complete set of available clinical data at the given time point. Each box plot represents the quartiles of the *auroc distribution across 1000 bootstrap samples, with whiskers extending up to 1.5 times the interquartile range. The median and mean are drawn as solid and dashed lines, respectively.
• Figure 5: Our HSI + clinical data models outperform widely used clinical biomarkers and scores for sepsis diagnosis and mortality prediction. Comparison of the *auroc for deep learning-based sepsis diagnosis (A) and mortality prediction (B) using *hsi data of the palm (HSI palm model, gold) and a combination of *hsi data and the entire set of clinical data available within one hour (left) and ten hours (right) from admission to the intensive care unit (HSI palm + clinical data model, bronze) against clinical biomarkers and scores (blue). For data available within one hour of *icu admission, the comparison includes *news, *crt, *sms, *qsofa score, and *vis. For data available within ten hours of admission, the comparison includes *crp, *pct, *sirs criteria, *sofa score, and *apache II score. Each box plot displays the quartiles of the *auroc distribution across 1000 bootstrap samples, with whiskers extending up to 1.5 times the interquartile range. The median and mean are represented by solid and dashed lines, respectively.