Fig. 1: Contemporary AI models for bacterial infection diagnostics.

a Conformal Multidimensional Prediction Of Sepsis Risk (COMPOSER) collects 40 clinical variables from EHRs, including dynamic variables such as vital signs and laboratory measurements, and demographic variables such as age and gender. This information is used for a weighted input layer that scales the value of a clinical variable dependent on the most recent measurement. This input layer is fed into a FFNN to reduce data dimensionality. The dimensionally reduced vector is fed into a conformal predictor that uses two conformal sets to quantify the conformity of new patient-level features to previously seen septic and non-septic examples. This enables the model to identify outliers not meeting the algorithm's conditions, assigning them to an indeterminate label class. If the data conforms, it is fed into a sepsis predictor (FFNN) that predicts the probability of sepsis from 0 to 1. b A 3D tomogram of a single cell or cell cluster is generated using 3D QPI tomography and assembly of 2D sinograms. The 3D tomogram is input into a CNN with four dense blocks containing 12, 24, 64, and 64 convolution operations which undergo batch normalization and average pooling between each block. The final output is a 19-dimensional vector that contains the conditional probability of the 3D RI tomogram being one of 19 BSI causing bacteria. c Model was trained on DRIAMS, which contains mass spectra and an AMR profile for each instance. A collected mass spectra is inputted into the model, which is then pre-processed and binned into 3 Da ranges from 2000 Da to 20,000 Da. These bins are vectorized, and inputted into the appropriate model (LR, lightBGM, or MLP) to generate a prediction of resistance to a certain antibiotic.