Skip to content
EN
Contact

Data science consultancy in The Hague

We test a predictive model using recent data and assess its performance for each group before anyone uses it to make decisions. For LUMC, we put our own models into practice.

Discuss your model Read the LUMC case study

LUMC’s own machine learning models have been put into practice

LUMC, also a partner of Health Campus Den Haag, developed several models to improve healthcare. We set up the infrastructure to put them into practice, with the results presented in a clear application.

Read the case study

Does our model still hold up for today’s neighbourhoods and patients in The Hague?

A model performs well on the dataset it was built with. But residents and patients differ between neighbourhoods, and how data is recorded and care is delivered changes. Testing it on recent data and for each group shows whether it holds up there too.

Book a meeting

A predictive model tested on the group it will be used for

Discuss your model

Start with the decision

We start with the decision that needs improving, such as which client gets an intake appointment first, and only then choose the data and model.

Tested across groups and over time

Beyond the overall score, we examine how the model performs by age group, neighbourhood or location, and on data collected after it was built.

A probability that holds up in practice

A predicted probability should occur just as often in practice. This is called calibration, and we test it separately from accuracy.

In use, not in a report

For LUMC, we are setting up the infrastructure to deploy its own models, with the results presented in a clear application.

We establish the legal basis and the question before training a model

  1. Question and legal basis

    Which decision needs to improve, who will use the outcome, and can the data be used for that purpose?

  2. Build and compare

    We build the model, test it on recent data and across groups, and compare it with the current way of working.

  3. Run and monitor

    The model runs where the user already works. We monitor whether the input data and scores shift.

Is a predictive model right for our question?

Discuss whether it’s a good fit

A good fit when

There’s a recurring decision, such as which case to review first or who should receive extra support, and the outcomes of previous cases have been recorded.

Not a good fit when

The outcome is rare or poorly recorded, or a fixed policy or guideline already determines the decision. In that case, better record-keeping or an overview would help more, and we’ll tell you so.

What determines the scope

The number of sources and how they’re linked, whether the data can leave the organisation, how much validation is needed, where the model runs and who manages it afterwards.

Can the data be used for this?

Health and social services data is often collected for another purpose. Your data protection officer or legal adviser decides whether it can be reused. We build the model to use as little personal data as possible.

Analysis first, a simple model or machine learning: which fits your question?

Start with an analysis

Choose this if it is not yet clear which decision needs improving. Analysing your existing data shows where the differences lie and whether a model would add value.

A simple model

Choose this when explainability matters, such as for decisions affecting residents or patients. Our advice: always build a regression model first as a benchmark; a more complex model must demonstrably outperform it.

Machine learning

Choose this when you have many data sources and features, and a simple model falls short. Read more about developing machine learning solutions.

For healthcare models between The Hague and Leiden, LUMC is our reference

Data Science Lab has worked with data and AI since 2016, from data platforms to applications and management, based in Amsterdam. We are a Databricks and Microsoft partner. We can meet at your premises, at ours or online.

Book an introductory meeting

Bring one decision and the data to the first meeting

Stef Stoelwinder

Stef Stoelwinder

Business Development

Stef is your first point of contact. In our first meeting, we’ll look at the decision that needs improving and the data behind it.

E-mail info@datasciencelab.nl

Phone +31 (0)20 3636 724

Which recurring decision would you like to support with a prediction?

Describe the decision, the data you have and who will work with it. We’ll reply by email or arrange an initial conversation.

Emailinfo@datasciencelab.nl Phone+31 (0)20 3636 724

Enter a valid email address.

Let us know how we can help you.

Frequently asked questions

It builds models using your own data to support recurring decisions: which patient or client needs extra attention, which report to handle first, or how much capacity is needed. This also includes text, such as classifying open-ended responses by topic; we built a topic classification model for WUA. These models are part of the AI applications we build and then manage. For public-sector organisations outside healthcare, we have also written about data and AI for education and housing associations.

LUMC developed several machine learning models to improve care. The aim was not just to develop them, but to implement them too. We set up the infrastructure, after which the models were implemented in practice. The results are presented in a clear application, making their impact visible in day-to-day care.

Read more in the LUMC case study.

Health Campus Den Haag, a collaboration involving LUMC, Gemeente Den Haag, Haaglanden Medisch Centrum, HagaZiekenhuis and GGD Haaglanden, among others, studies health differences between residents. Partners share data in a secure CBS environment.

There, CBS links sources such as GP data from the ELAN network and municipal WMO data at the individual level, using a pseudonymised key. Only aggregated results may leave the environment, and only for scientific research.

The Guideline for Quality of AI in Healthcare, drawn up in December 2021 by a task force with support from VWS, divides the work into six phases: data collection and management, development, validation, software development, assessment of added value in practice, and implementation including education. It has no legal status.

At a minimum, test a model on data it did not see during development, such as data from a later period or a different location.

A model can rank patients or records correctly while still estimating probabilities incorrectly. If it consistently predicts probabilities that are too high, too many people receive an extra check-up or home visit, and scheduling becomes unmanageable.

You test calibration by comparing predicted probabilities with how often the outcome actually occurred in each group. If they diverge in a new period, you adjust the model.

Sometimes. Point 5 of Annex III covers AI used by public authorities to assess whether someone is entitled to essential public benefits and services, including healthcare, as well as systems for triaging patients in emergency care. According to the European Commission, those rules apply from 2 December 2027.

Your legal or compliance team can assess whether your model falls within its scope; we can help identify the relevant systems. See AI Act.

By continuing to monitor its inputs and outcomes. A new registration system, a different protocol or a change in who seeks help alters the data the model relies on. Once the actual outcomes are known, you compare them with the predictions.

Read how we set this up under MLOps; ongoing maintenance is part of management and continuous improvement.

If you have not yet identified a decision that needs improving, or if you first want to understand where AI fits in your organisation and how to use it responsibly, start with AI consultant in The Hague.

If your organisation is in the Utrecht region, see data science consultancy in Utrecht.

Want to explore the model, the technology or the advice further?

View all services