We present a system for data-driven therapy decision support based on techniques from the field of recommender systems. Two methods for therapy recommendation, namely,
The large volume of daily captured data in healthcare institutions and out-of-hospital settings opens up new perspectives for healthcare. Due to the amount of that data, its high dimensionality and complex interdependencies within the data, an efficient integration of the available information is only possible using technical aids. In this regard, data-driven clinical decision support systems (CDSS) can be expected to take a major role in future healthcare. Generally, CDSS are designated to assist physicians or other health professionals during clinical decision-making. CDSS are demanded to be integrated into the clinical workflow and to provide decision support at time and location of care [
While most works related to CDSS deal with diagnosis decision support [
Within this contribution, we present a system for therapy decision support based on techniques from the field of recommender systems which originates from E-commerce and has developed considerably over the last years. Recommender systems are able to overcome the aforementioned limitations of traditional data-mining and machine-learning techniques, which render suchlike systems an interesting alternative for therapy decision support. In medicine, however, the application of recommender systems is rather limited. In [
In general, CDSS can be classified into knowledge-based and data-driven approaches having both advantages and suffering from disadvantages. Knowledge-based systems on the one hand usually rely on manually encoded rule-based expert knowledge (if-then rules) to infer decision support. Applied rules typically represent clinical guidelines and best practice rules providing a reliable decision basis [
Data-driven approaches on the other hand apply methods from data-mining and machine-learning to automatically extract knowledge from clinical data, facilitating more individual recommendations, learning from past experience, and revealing unknown patterns in the available data [
Recommender system technologies date back to the nineties [
Over the years, the field of recommender systems has evolved considerably yielding extremely sophisticated and specialized methods depending on domain, purpose, and personalization level [
A basic taxonomy of recommendation algorithms differentiates between content-based [
Concerning data-driven therapy or treatment recommendation in general, some scientific works were proposed, ranging from approaches based on majority voting [
In this work, we transfer the idea of collaborative filtering to the domain of CDSS. We present a recommender system which aims at predicting the adequacy of different therapy options for a given patient at a given time. To that end, two methodologies for therapy adequacy estimation, a
In this work, different recommender system approaches are developed and evaluated based on excerpts from health records provided by the Clinic and Polyclinic for Dermatology, University Hospital Dresden. The data consists of
The different attributes making up the data are of various levels of measurement ranging from dichotomous to ratio-scaled attributes. Moreover, in spite of data padding in cases where information was missing but could be assumed to be constant over consultations, availability of certain attributes is very limited. As a consequence, the resulting data matrices are characterized by inhomogeneity and sparsity. Patient attributes and therapy information are summarized in Tables
Patient describing attributes.
Attribute | Scale | Range | Availability % |
---|---|---|---|
|
|||
Year of birth | Interval | 1931–1998 | 100 |
Gender | Nominal | 1, 2 | 100 |
Weight | Ratio | 50–165 | 51.40 |
Size | Ratio | 99–204 | 35.73 |
Family status | Dichotomous | 0, 1 | 53.02 |
Planned child | Nominal | 1, 2, 3 | 8.01 |
Year of first diagnosis | Interval | 1950–2014 | 89.74 |
Type of psoriasis | Nominal | 1, 2, 3, 4, 5, 6 | 100 |
Family anamnesis | Ordinal | 1, 2, 3 | 50.95 |
|
|||
|
|||
Comorbidity | Nominal | 1, 2, 3, ..., 34 | — |
Status | Ordinal | 1, 2, 3 | 100 |
Under treatment | Dichotomous | 0, 1 | 100 |
Disease-free | Dichotomous | 0, 1 | 100 |
|
|||
|
|||
PASI score | Ratio | 0–43 | 70.57 |
Self-assessment severity | Ordinal | 1, 2, 3, 4, 5 | 9.45 |
Development face | Ordinal | 1, 2, 3 | 6.84 |
Development feet | Ordinal | 1, 2, 3 | 9.81 |
Development nails | Ordinal | 1, 2, 3 | 20.97 |
Development hands | Ordinal | 1, 2, 3 | 12.33 |
Treatment contentedness | Ratio | 0–10 | 10.62 |
Therapy describing attributes.
Attribute | Scale | Range | Availability % |
---|---|---|---|
Systemic |
Nominal | 1, 2, 3, ..., 15 | — |
Effectiveness | Ordinal | 1, 2, 3 | 23.67 |
ΔPASI | Ratio | −27–18 | 42 |
Adverse effect | Dichotomous | 0, 1 | 100 |
|
|||
Systemic therapy | Nominal | 1, 2, 3, ..., 15 | — |
Effectiveness | Ordinal | 1, 2, 3 | 98.43 |
ΔPASI | Ratio | −27–18 | 42 |
Adverse effect | Dichotomous | 0, 1 | 100 |
Previous treatment is the collection of all relevant therapies applied to a patient up to the consultation under consideration, whereas in the current treatment database, all therapies are collected which were applied within the last two weeks preceding the respective consultation. Even though there is information on both local and systemic therapies available, this study focuses on recommending the most effective systemic therapy out of
The algorithms described in the following aim at recommending the potentially most effective systemic therapy for a given patient and consultation. The collaborative filtering idea is transferred to the therapy recommendation domain, considering therapies as items and therapy response as a user’s preference. For representing therapy response, effectiveness, ΔPASI, and absence of adverse effects are incorporated.
In a preceding prediction step, individual therapy outcome is estimated for all available therapies that have not yet been applied to the patient. The outcome estimate is computed based on the therapy response of the nearest neighbors to the consultation under consideration. At this stage, similarity computation between consultation representations plays an essential role. The two recommender approaches proposed in this work differ in the information used to represent consultations. The applied
In typical recommender system applications, data reflecting a user’s preferences is collected from both explicit and implicit input. Where explicit expression of preference usually is provided as item ratings or votes on items, implicit information can be derived from clicked items, items being part of the shopping basket, or visited pages, respectively. Here, the preference to a therapy is derived from the therapy response. The mathematical quantification of therapy response, in the following denoted as
Sigmoid function mapping the ΔPASI score to the
Both proposed methods,
Figure
Affinity matrix for 25 randomly selected consultations.
One approach to address the trust that can be placed in the similarity to a neighboring consultation, depending on the available information, is significance weighting. In case of the
Finally, the computed overall similarity is normalized with the sum of weights
Significance weighting in case of the
To generate an affinity estimate on appropriate therapies for a consultation under consideration, various methods for computing aggregates of neighboring consultations’ therapy response, that is, affinity, are compared. Here, an affinity estimate
For all introduced approaches, the summations are performed over the
As stated beforehand, both proposed recommender approaches have their strengths and weaknesses. The idea of building a recommender ensemble is to generate an overall recommendation which combines both approaches while compensating for the individual recommender engines’ drawbacks. Fusing decisions in machine-learning applications, denoted as ensemble learning, have shown to be capable of outperforming basic algorithms [
In this work, two different evaluation metrics are considered. On the one hand, the individual recommender engine or system yields to predict the response to specific therapies. If the prediction meets the real therapy response, the system can provide the medical practitioner with a reliable support for his decision-making based on the estimation. To quantify the difference between estimated response and real response, the root mean squared error (RMSE) for a specific consultation is computed between provided affinity entries and predictions. RMSE reflects the rating error in the same value domain as the actual affinity measure with large errors having more impact [
On the other hand,
Outcome-driven evaluation definitions.
Good outcome | Bad outcome | |
---|---|---|
Recommendations compliant |
|
|
Recommendations not compliant |
|
|
The outcome-driven precision describes the ratio of all therapies recommended by the system for a consultation
In the following, the three affinity estimation approaches introduced in Section 3.2.4 are compared for both the
RMSE and precision at 3
RMSE and precision at 3
RMSE and precision at 3
Relative number of consultations for which affinity estimation RMSE can be computed depending on
Regarding recommendation precision, the ground truth is obtained from all consultations having one or more therapies which showed good response according to the definition described in Section 3.3. As a result, precision can only be computed for 67.24% of all consultations in the database for the
In both nonnormalized
As stated beforehand, for the
In this work, the application of recommender system algorithms in the context of therapy decision support was studied. Even though there is an extensive impact of recommender systems in other domains, application of suchlike approaches in healthcare are—to the best of our knowledge—still rare to date. Dependent on the data employed for determining similarity between consultations and therapy outcome estimation, two approaches were compared. For both algorithms, a
The
Applying patient describing data for similarity computation, that is, the
Further on, beyond the presented comparison of different recommender-based approaches, a comparison of the proposed methods to alternative machine-learning algorithms for generating therapy recommendations, particularly model-based approaches, would be of high interest. However, one of the major reasons to apply recommender methods are their capability to handle heterogeneous and sparse data. The application of typical model-based approaches, in turn, is difficult as structure and characteristics of the clinical data at hand, that is, its high degree of heterogeneity and sparsity, would require extensive feature transformation and preparation (handling of missing data, transformation of non-interval-scaled data). As suchlike preparation is complex and will heavily impact the performance of machine-learning algorithms, the usage of such techniques and their comparison to the recommender approaches exceeds the extent of this work. However, future works will address this issue considering the presented clinical data but also using other data in order to yield a comparative assessment of the proposed methods.
This research article is an extended version of the conference paper presented at the 2016 IEEE 18th International Conference on e-Health Networking, Applications and Services (Healthcom).
The authors declare that there is no conflict of interest regarding the publication of this paper.
The work is part of the project