Early detection of hepatocellular carcinoma in high risk populations using MALDI-TOF mass spectrometry
Inventors
Röder, Joanna • Oliveira, Carlos • Grigorieva, Julia • Röder, Heinrich • Mahalingam, Devalingam
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-10037874-B2
Publication Date
2018-07-31
Expiration Date
Abstract
Hepatocellular carcinoma (HCC) is detected in a patient with liver disease. Mass spectrometry data from a blood-based sample from the patient is compared to a reference set of mass-spectrometry data from a multitude of other patients with liver disease, including patients with and without HCC, in a general purpose computer configured as a classifier. The classifier generates a class label, such as HCC or No HCC, for the test sample. A laboratory system for early detection of HCC in patients with liver disease is also disclosed. Alternative testing strategies using AFP measurement and a reference set for classification in the form of class-labeled mass spectral data from blood-based samples of lung cancer patients are also described, including multi-stage testing.
Core Innovation
The invention provides a method of generating a classifier for detecting hepatocellular carcinoma (HCC) in a patient with liver disease by using MALDI-TOF mass spectrometry on blood-based samples from a multitude of patients. Mass spectrometry data from patients with liver disease are stored as a development set including feature values at a multitude of mass spectral features, and classification labels are assigned to each member of the development set based on whether the associated patient had HCC.
A multitude of mini-classifiers are constructed using one or more of the feature values, and the performance of the mini-classifiers on the training set is filtered by retaining only mini-classifiers that meet a performance threshold. The retained mini-classifiers are then combined to generate a master classifier using a regularized combination method, and the master classifier performance is evaluated on the test set.
The method repeats the training-test separation, mini-classifier construction, filtering, master classifier generation, and test-set evaluation for many different realizations of the separation into training and test sets. A final classifier is defined from the master classifiers resulting from the repeated iterations, and the described framework is used to detect HCC/No HCC using the integrated intensity values at predefined mass spectral features and can optionally include serum AFP-related features.
Claims Coverage
Independent claim clm-00001 covers a training-and-selection framework with 8 inventive features: constructing and filtering mini-classifiers from mass spectral feature values, combining retained mini-classifiers into a master classifier via a regularized combination method, evaluating on test sets, repeating across train/test splits, and defining a final classifier from the resulting master classifiers. Dependent claims further refine the framework with AFP-related exclusions or AFP expression features, a healthy-patient accuracy threshold, K-nearest neighbor mini-classifier settings, and a minimum feature-count constraint.
Mini-classifier construction from mass spectral feature values with filtered retention
Constructing a multitude of mini-classifiers using one or more of the feature values, filtering the performance of the mini-classifiers operating on the training set, and retaining only those mini-classifiers that meet a performance threshold.
Master classifier generation by regularized combination of retained mini-classifiers
Generating a master classifier by combining the filtered mini-classifiers using a regularized combination method.
Repeated train/test realizations to define a final classifier
Repeating the steps of separating into training and test sets, constructing mini-classifiers, filtering, generating the master classifier, and evaluating on the test set for many different realizations of the separation, and defining a final classifier from the master classifiers resulting from the repeated iterations.
AFP-based exclusion of high AFP expression patients in the development set
Defining the development set to exclude patients with high AFP expression levels.
AFP expression level as a feature value
Using feature values that include an AFP expression level measured in the blood-based samples of a development set.
Healthy-patient accuracy thresholding for mini-classifier retention
Testing mini-classifiers on spectra of healthy patients during the filtering step and retaining only if healthy patient classification accuracy exceeds a pre-determined limit.
K-nearest neighbor mini-classifiers with discrete K values
Implementing the mini-classifiers as a K-nearest neighbor algorithm with K equal to 5, 7, 9, or 11.
Constraint to at least 50 listed features for mini-classifiers
Using mini-classifiers that use at least 50 features from the set of features listed in Tables 1, 2, or 4.
Across the independent claim and its dependent refinements, the claim coverage emphasizes selecting high-performing mini-classifiers from MALDI-TOF mass spectral feature values using performance thresholds, combining retained mini-classifiers into a regularized master classifier, and defining a final classifier from repeated train/test evaluations, with optional AFP-related handling, healthy-patient accuracy filtering, specified K-nearest neighbor settings, and constrained feature-set size tied to listed tables.
Stated Advantages
Improved detection performance is supported by performance evaluation of the master classifier on test sets and by selection of mini-classifiers that meet performance thresholds, including filtering using healthy-patient accuracy thresholds.
The classifier-generation framework supports robust final classifier definition by repeating across many different realizations of the separation of the development set into training and test sets.
Documented Applications
Early detection of hepatocellular carcinoma (HCC) in a patient with liver disease in a high-risk liver-disease population using blood-based samples and MALDI-TOF mass spectrometry with a computer-implemented HCC/No HCC classifier output.
Laboratory testing system for MALDI acquisition, preprocessing, feature extraction, and CMC/D classification to output HCC/No HCC labels.
Multi-stage testing strategies that leverage VeriStrat-like classification and/or AFP thresholds before (or alongside) the MALDI HCC/No HCC classifier.
Interested in licensing this patent?