This paper presents a solution to the challenge of cross-device model generalization in blood-based infrared spectroscopy. As infrared spectroscopy becomes increasingly popular for analyzing human blood, ensuring that machine learning models trained on one device can be effectively transferred to others is essential. However, variations in device characteristics often reduce model performance when applied across different devices. To address this issue, we propose a straightforward domain adaptation method based on data augmentation incorporating device-specific differences. By expanding the training data to include a broader range of nuances, our approach enhances the model’s ability to adapt to the unique characteristics of various devices. We validate the effectiveness of our method through experimental testing on two Fourier-Transform Infrared (FTIR) spectroscopy devices from different research laboratories, demonstrating improved prediction accuracy and reliability.

Human biofluids are valuable indicators of physiological states, and recent advances in molecular profiling technologies offer significant promise for improving clinical diagnostics. In this study, we evaluate the potential of laser-based electric-field molecular fingerprinting as a novel tool for in vitro diagnostics. In a proof-of-concept clinical study involving 2,533 participants, we conducted randomized measurement campaigns to spectroscopically profile bulk venous blood plasma across four cancer types: lung, prostate, breast, and bladder. Using machine learning, we identified infrared spectral signatures associated with therapy-naïve cancer states, distinguishing them from matched control individuals. For lung cancer, we achieved a cross-validated area under the receiver operating characteristic curve (ROC AUC) of 0.88, while ROC AUC values ranged from 0.68 to 0.69 for prostate, breast, and bladder cancers. In an independent held-out test set—designed to reflect experimental conditions different from those used during model training—we achieved a ROC AUC of 0.81 for lung cancer detection. These results demonstrate that electric-field molecular fingerprinting provides a robust and scalable framework for disease phenotyping, with potential for broad application in real-world diagnostic settings.

At the heart of the research is a denoising autoencoder — a type of neural network designed to extract meaningful patterns from complex data. Applied to Fourier Transform Infrared (FTIR) spectroscopy, the model employs a custom loss function and a bottleneck architecture to strip away noise while preserving critical molecular information. The result: a 2.6-percentage-point boost in lung cancer detection accuracy in a case-control study. Beyond improved performance, the learned latent space also highlights spectral features linked to disease presence, paving the way for more insightful and interpretable diagnostic tools. This work marks the first-ever application of modern AI techniques to infrared molecular fingerprints from blood samples collected in the framework of attoworld.

About eight years ago, the Broadband Infrared Diagnostics (BIRD) research group and the clinical study team,Lasers4Life at attoworld set out to collaborate with Prof. Dr. Jürgen Behr and the medical professionals from the Asklepios Clinic in München-Gauting and the Department of Medicine V of LMU University Hospital in Munich. Together, they initiated an investigation into the intricacies of chemical footprints left by lung tumors in the blood and how these profiles change over time. The study tracked individuals from the time of their lung cancer diagnosis throughout their treatment. Given that lung cancer is unfortunately still diagnosed at an advanced stage only, there is a critical unmet need for minimally invasive diagnostic tests that enable earlier detection. In the current study, the BIRD team applied Fourier-transform infrared (FTIR) spectroscopy as a molecular fingerprinting technique to profile blood samples from 160 lung cancer patients. Meanwhile, the CMF Data Science team analyzed the acquired molecular profiles and applied statistical and machine learning methods to assess patient survival. Their findings revealed that the infrared fingerprints correlated with lung cancer survival in a manner comparable to traditional tumor staging and blood biomarkers. Moreover, a larger case-control comparison of 501 individuals demonstrated an association between these infrared fingerprints and lung cancer progression.

The generation of laser pulses with controlled optical waveforms, and their measurement, lie at the heart of both time-domain and frequency-domain precision metrology. Here, we obtain midinfrared waves via intra-pulse difference-frequency generation (IPDFG) driven by 16-femtosecond near-infrared pulses, and characterise the jitter of sub-cycle fractions of these waves relative to the gate pulses using electro-optic sampling (EOS). We demonstrate sub-attosecond temporal jitter at individual zero-crossings and sub-0.1%-level relative amplitude fluctuations in the 10-kHz–0.625- MHz band. Chirping the nearly-octave-spanning mid-infrared pulses uncovers wavelength-dependent attosecond-scale waveform jitter. Our study validates EOS as a broadband (both in the radiofrequency and the optical domains), highly sensitive measurement technique for the jitter dynamics of optical waveforms. This sensitivity reveals outstanding stability of the waveforms obtained via IPDFG and EOS, directly benefiting precision measurements including linear and nonlinear (infrared) fieldresolved spectroscopy. Furthermore, these results form the basis toward EOS-based active waveform stabilisation and sub-attosecond multi-oscillator synchronisation/delay tracking.