PROBLEM: Building preconfigured machine learning based solutions on Nebbiolotech platform for anomaly detection and predictive maintenance
SOLUTION: Our preconfigured UDF chains uses artificial intelligence to detect anomalies to prevent breakdown of your services by performing appropriate analytics on your streaming data at the edge itself so that you are immediately alerted when you need to take remedial action and know what action will be appropriate at that time. Our anomaly detection methods provide a ready solution generating alerts at different severity levels so that effective action can be taken by the production managers. We also predict the future performance of different machine parameters based on past data.
Anomaly Detection and Predictive Analytics
PROBLEM: We were engaged by Nebbiolo to assist with a problem from a large Automobile manufacturer with Robots on the shop floor to analyze vast amounts of sensor data to trap anomalies. A history of anomalous events was not available to facilitate detection.
SOLUTION: Analytics Plus captured and analyzed sensor data from Kuka Robots using sensors deployed by Nebbiolo. We used unsupervised data cleaning algorithms based on clustering (K-means/Dbscan) to detect anomalies and eliminate sensor error or missing data. Prediction algorithms like HoltWinters, EWMA, ARIMA were used to predict future movement of data under normal conditions and deviations from the norm was used to predict anomalies. An App was built on the Nebbiolo platform allowing users to use/ set up pipelines to analyze data from sensors.
BENEFIT: The solution developed on test lab robots was able to capture genuine anomalies in real time and predict future robot arm movements based on past data. This predicts anomalies in advance and thus allows predictive maintenance and root cause analysis to be done on shop floors.
Tools Used: Influxdb, Kapacitor, Grafana, Python.
UrjaSenseTM - Energy Management Solution in Industrial and Home Settings
UrjaSenseTM is an energy management and advisory software application for home owners. Its purpose is to use artificial intelligence and machine learning based algorithms to provide real-time insights into energy consumption in the overall home and each appliance in the home using electrical energy. Thus, providing evidence based advisory to home owners where they can make changes to their energy consumption behavior and time of usage of electricity so as to minimize their current electricity cost. This application is built on infrastructure that currently resides on Nebbiolo’s Fog Node and thus can be ported to it when the production version of the UrjaSense is available.
We use artificial intelligence based solutions to help you detect anomalies to prevent breakdown of your services by performing appropriate analytics on your streaming data at the edge itself so that you are immediately alerted when you need to take remedial action and know what action will be appropriate at that time. Our machine learning based anomaly detection methods provide a ready solution generating alerts at different severity levels so that effective action can be taken by the production managers. We also predict the future performance of different machine parameters based on past data.
One of the best teams in the country on claims based analytics, we have experience in working with the entire National Claims Data, and additional external data sources to allow our customers take important business decisions using our automated artificial intelligence based analysis. Some of the ways we have helped our clients are as follows
1. Physician Network and Referral Analysis:
Hospital and ACO management need to understand who are referring to their physicians and what proportion of their clients are being leaked to competitors within their market area. We have developed Loyalty and Volume Score for physicians in the network within the defined market of a client organization and helped identify the most likely physician’s practices that can be acquired to improve business.
2. Entity Resolution:
With increasing volume and velocity of health data, inference across networks and semantic relationships between the health organizations becomes a greater challenge leading to duplicate entities and linkage issue. We use a similarity quotient approach to generate a single entity key by standardizing the organization names and resolving different keys for the same organization using similarities of locations
3. Targeted Campaign for Hospitals:
Propensity Models for largest hospital chains in the US to determine prospective consumers for health products related to specific service lines using claims data for patients and survey data for potential consumers in the market area. Models were developed for more than 65 different diseases with patient models having more than 75% response rate and consumer models with greater than 65% success rate.
PROBLEM: Developing an Interactive Alert System through the Case Manager Dashboard to Identify and Respond to Patients Requiring Immediate Attention
Diabetes Patients at a reputed hospital were monitored through the IntelliH Remote Health Monitoring System. Case Managers handling many patients needed to focus on the patients that required immediate attention and an algorithm needed to be formulated to identify such patients with priority levels for this purpose.
SOLUTION: Anytime the patients experienced Blood Glucose excursions and other vital parameters satisfied certain conditions identified by our artificial intelligence based algorithm developed in consultation with their respective Physicians, alerts were sent to the Care Managers so that they could take appropriate action.
BENEFIT: Patients have commented on how timely intervention on the part of their Care Managers helped them avert emergency situations.
Development Environment: Visual Studio 2013, .Net Framework 4.5, Microsoft SQL Server 2008 R2 ; Coding Tools: Asp.Net, C#/VB.Net language, LINQ, Web API. Database: Microsoft SQL Server 2008 R2; External Libraries Used: Clickatell SMS Service, Google GCM (Google Cloud Messaging), .iOS .p12 Certificate
PROBLEM: Determining Severity of Patient Tweets Using Natural Language Processing For Appropriate Action by Case Managers.
Diabetes Patients participating in a study were asked to report their problems and daily symptoms encountered as well as any important information they wanted to communicate to their doctors in daily tweets. The problem was to assess severity of the tweets so that appropriate timely action could be taken.
SOLUTION: AnalyticsPlus analyzed these text messages to identify emotions expressed by the patients, the overall polarity of the note, with keyword extraction using the Snomed database to classify these notes into different priority levels so that Case Managers could take appropriate action based on severity.
BENEFIT: This process allows Case Managers to spend more times on patients experiencing more severe symptoms.
Tools Used R, Snomed Database
PROBLEM: Meta-Analysis to determine Most Effective Treatments and Dosages by Combining Results from Different Studies
When Comparing Different Competing Treatments at different dosages it is often not possible to do this in only one study because of ethical, monetary and other constraints. The task here is used to combine results from several studies which meet certain criteria and compares some but not all of these treatments at some of these dosages and make meaningful interpretations for all of the treatments at all of these dosages.
SOLUTION:We compared treatments for our client using Models and links like Conditional Binomial Likelihood, with probit link, random and fixed effects model, binomial likelihood, logit link, random & fxed effect models and created network diagrams and rankograms with best probability calculation. For this purpose we did pairwise meta-analysis and indirect treatment comparison. Both frequentist and Bayesian framework comparisons were made and patient characteristics were also analyzed.
BENEFIT: The cost of new clinical trials to test all possible combinations of treatments and dosages were avoided by our clients and our models were used to determine the effective dosages of treatments and identify most effective treatment and dosages for different levels of diseases.
Tools: R, Stata, WinBugs
PROBLEM: Automated detection of Arrhythmias from ECG Reports
Since both Doctors and patients are busy it is not always possible for the patient to visit the doctor, have an ECG done and get it checked manually. There is a need for reliable automated analysis of ECG Reports for patients who are away from their doctors so that timely interventions can be made to prevent readmissions and Emergency Room visits.
SOLUTION: Analytics Plus has developed an algorithm, working with publicly available ECG records to detect the presence of bradycardia, tachycardia, atrial fibrillation and atrial flutter through machine learning algorithms. This algorithm can be integrated with a mobile app that allows the patients to take an ECG reading from their mobile which is analyzed by this computer generated algorithm and forwarded to their physician. The report is prioritized whenever the algorithm detects abnormalities so that the report is personally reviewed by the Physician and recommendations sent immediately back to the patient.
BENEFIT: Patients can receive recommendations from the doctor even when they are travelling and possible readmissions and emergency room visits can be avoided.
Tools Used: R, Tableau
PROBLEM: Workers Compensation Claims Triaging- Identifying Expensive Claims Early in their Life Cycle
When the claim is first generated, only minimum information is available. The assessors allocate funds based on their experience and often opportunities are lost where escalation could have been prevented by taking appropriate steps at the right time. Again, when more information becomes available at a later date the best use of this information is important to make appropriate adjustment to allocated resources.
SOLUTION: AnalyticsPlus Developed a scoring algorithm to identify expensive claims at the 0-th day or the day they are first reported. Proprietary algorithms were used to create a Severity Index that assessed the likelihood of the claim escalating rapidly by combining claim related injury related data available on the claim and each claim was scored at 0-daya. This score was updated using another model making use of variables available at this stage at the 30 day stage.
BENEFIT: The High Predictive accuracy of these models allowed our clients to assign appropriate resources to these claims early on in their lifecycle helping them reduce the ultimate costs on these claims.
Tools Used: R
Quality Assessment, Cleansing, and Reporting of Medical Devices and Health Sensor
As a data analytics company, AnalyticsPlus has extensive experience in health care data quality assessment and cleansing. While humanly collected data has issues related to missing, incorrect data entries of different varieties, data duplication, and many others , all kinds of data in health whether sourced from EMR, claims, or patient surveys require significant quality assessment and cleansing before they can be used for analysis, reporting and visualization. The problem is compounded when multiple sources and types of data are merged together to meet the statistical or machine—learning based analytics or predictive modeling. As the number of health sensor and medical devices becoming available to patients are increasing exponentially, the need for data quality assessment, cleansing and validation is also increasing.
We are developing a “clearing house” of 100s of health sensor data through a systematic quality assessment, cleansing, and reporting platform, called QualitySense, that can be integrated with the Electronic Medical Record of health care facilities. Thus, all the data coming from these personal-sensors can be vetted and a seal of “AnalyticsPlus Approval” can be designated for those devices. Unlike one time FDA approval process, QualitySense system will be processing data coming from these devices on a real time basis and validating them or flagging them by comparing within device variability, calibration data, and inter-device variability. Thus, the hospital management and their care providers have a constant visibility and can see which devices/brands are more reliable than others and encourage their patients to use those accordingly. A layered or sequential pipeline approach to data quality assessment and cleansing has been developed using streaming data analytics infrastructure.
QualitySense will be our proprietary data quality tool to be used universally. It will enable cleansing and monitoring of data quality on an ongoing basis, helping to turn input data into trusted information. This Application will use appropriate business rules at data entry, during continuous streaming, and at the end point, and scale to cleanse millions of rows of static or dynamic data to provide trusted information for Data Analysis.