Transition to Production
For use in production, interfaces, permissions, monitoring, versioning, and responsibilities must also be clarified.
- Clearly Defined Use Case
The business question, target users, and success metrics are clearly defined. - Data Readiness
The required data is available in sufficient quality for training and ongoing use. - Validated Model
Model quality and limitations have been verified and documented based on established criteria. - Integrated system
Applications can incorporate and process model results.
The Right AI Approach
When Data Science & Machine Learning Is the Right Approach
Not every AI problem requires generative AI. Depending on the available data and the objective, data science and machine learning are the right approach—especially forforecasting, scoring, classification, and the detection of patterns in structured data.
Machine Learning Consulting
Our Machine Learning Consulting Services
Our ML consulting combines business assessment, solution design, model development, and the implementation of data-driven solutions. Depending on the starting point and level of maturity, we provide support ranging from use-case evaluation, data preparation, model training and validation, to technical integration, rollout, and scaling
Real-World Examples
Experience with Data Science and Machine Learning
HMS develops machine learning solutions and integrates them into existing processes and systems. The following projects demonstrate how we solve specific business challenges using data science and machine learning and put models into production.
Covariate Selection Using Machine Learning for Clinical Models
HMS developed an interactive application for a global pharmaceutical company to support ML-based covariate selection. It provides a structured approach to selecting covariates and helps assess their relevance for pharmacometric models.
AI in Clinical Trials: Machine Learning for the Analysis of Clinical Data
Machine learning can complement traditional statistical analysis of clinical data. HMS supported Novartis in exploring its potential to predict tumor control and safety outcomes and to identify relevant baseline factors.
A Lab in Your Pocket: Data-Driven Decisions
HMS developed a modular software solution for trinamiX consisting of a mobile app, a cloud backend, and a web-based customer portal. The solution enables real-time transfer and analysis of sensor data from an infrared detector.
Benefit from our experience
What Sets HMS Apart in Machine Learning Projects
At HMS, data science and software engineering go hand in hand. Even during model development, we take into account how the model will later be integrated into applications and processes, operated, and further developed.
- End-to-end implementation from the model to the production system
- Data science and software engineering in the same project team
- Integration into existing applications, data platforms, and processes
- Experience with complex and regulated environments
Machine learning is particularly well-suited when decisions are regularly made based on data, when processes need to scale, or when complex relationships between influencing factors need to be identified.
Machine learning is less suitable when:
- there is very little data available or the data is flawed
- simple rules are sufficient
- decisions are made rarely and manually
The key factor is whether the solution aligns with the use case from both a technical and business perspective. In some cases, reporting tools and tailored systems for data preparation and implementing business logic may be sufficient as an alternative to a machine learning project.
Generative AI creates new content such as text, images, or code and is particularly well-suited for dialogue-oriented applications.
Machine learning identifies patterns in data and generates predictions, scores, or decision models. The data can be incorporated into the analysis as structured tables, as well as image or audio files.
For clearly defined evaluation and forecasting tasks, specialized machine learning models are often:
- more accurate
- more stable
- more reproducible
- more cost-effective to operate
However, the right choice always depends on the specific use case.
The amount of data required depends on the use case and the business question. Specialized methods can also be used with smaller data sets.
In any case, the following are crucial:
- Data quality
- Structure
- Subject-matter relevance
It is not “more data” but rather suitable and consistent data that is crucial. We systematically review the data set for quality, completeness, and relevance and predictive value before a model is deployed in production.
Typical phases include:
- Use-Case Validation and Goal Definition
- Data preparation and exploratory analysis
- Model selection, model training, and domain-specific validation
- Integration into existing systems
- Monitoring and further development
The goal is to create a solution that can be put into productive use and goes beyond one-time analyses.
A model only becomes truly useful once it is reliably integrated into existing applications and processes.
This includes:
- Interfaces to the relevant data sources and applications
- Version control for data, models, and code
- Monitoring of model quality and data changes
- Automated training and deployment processes
- Defined approval processes and operational responsibilities
Even after deployment, an ML model remains part of a running system: data changes, and models must be monitored and, if necessary, retrained or adjusted.
