Risk adjustment is an important stabilization program in the health insurance market that aims to reduce incentives for avoiding costly patients. Without risk adjustment, plans may have incentives to enroll healthier patients and to avoid sick patients (Schone and Brown, 2013; Ellis et al., 2018; Hileman et al., 2016). Under risk adjustment, plans would receive higher payments for patients predicted to cost more (e.g., patients with multiple chronic conditions) and less for patients predicted to cost less (e.g., healthy patients). Hence, accurate risk adjustment models not only counteract avoidance of patients with greater health care needs but also provide an opportunity to provide care efficiently and generate profits by attracting the less healthy patients (Schone and Brown, 2013).
Risk adjustment often begins with risk assessment, which assigns each patient a risk score that measures how costly that patient is expected to be (Hileman et al., 2016). To calculate the risk scores, a formula is developed that predicts the cost as a function of patients’ attributes such as demographic characteristics, prior expenditure, medical conditions extracted from diagnostic codes or medications, or self-reported health status (Chang and Weiner, 2010; Schone and Brown, 2013). Typically, risk adjustment models are estimated using classical linear regression (Rose, 2016; Chang and Weiner, 2010), which does not fully exploit the information in the data such as interactions and non-linear relationships between variables (Rose, 2016). In addition, although an increasing number of variables have been included in the models, they are unlikely to capture all of the factors that would affect expected costs (Schone and Brown, 2013). Considering the current performance of popular risk adjustment models ( ranged from 0.15 to 0.17) (Schone and Brown, 2013), there is a substantial potential for performance improvement in risk adjustment models. Besides, developing risk adjustment models is a complicated process (Ellis et al., 2018) which requires substantial feature engineering such as domain expert knowledge and data preprocessing.
In this paper, we propose to use semantic embeddings (Le and Mikolov, 2014; Mikolov et al., 2013b; Mikolov et al., 2013a) to extract patient representations for risk adjustment. Our work makes the following contributions:
We propose to use an established, easy-to-implement embedding algorithm to learn generic patient-level representations in claims data without depending on expert medical knowledge and heavy data preprocessing.
We apply the learned representations to predicting prospective risk scores and demonstrate the superior performance of the proposed semantic embeddings compared to a commercial risk adjustment model.
We use both linear and non-linear machine learning algorithms for predicting prospective risk scores and show the performance improvement of non-linear algorithms.
2. Related Work
Early statistical language models typically used schemes, such as one-hot encoding, that represented words as discrete, independent units. This sparse representation was inefficient to store and process. It also hampered the ability of these models to generalize patterns learned from one set of words to other combinations of words(Beam et al., 2018; Bengio et al., 2003).
Distributed representations of words solve both of these problems. First, they more efficiently represent words as dense vectors. Second, they can encode semantic and syntactic similarities and differences among words by arranging similar words near each other in the embedded vector space (i.e., semantic embeddings) and different words far apart. For example, cat and dog might be placed near each other, whereas both might be far from blue or write. Based on the similarity between cat and dog, a language model that is trained on word vectors and has encountered the sentence ”The cat runs.” might more readily generate the grammatically and semantically sensible sentence ”The dog runs.” (Bengio et al., 2003; Mikolov et al., 2013a)
Historically, the major obstacle to using high-quality word vectors was creating them within a feasible computational time. Mikolov et al. (Mikolov et al., 2013a; Mikolov et al., 2013b) largely overcame this obstacle by proposing the word2vec models, the continuous bag-of-words (CBOW) model and the continuous skip-gram model. Both models work by repeatedly selecting a target word and a specified number of surrounding context words. The CBOW model trains to predict the target word based on the context words’ vectors, while the skip-gram model predicts the context words from the target word’s vector. As successive sets of target and context words are used for training, the embedded word vectors are adjusted to minimize the prediction error so that high-quality embeddings are learned.
Le and Mikolov (Le and Mikolov, 2014) extended the word2vec models to groups of words, including sentences, paragraphs, and entire documents. In their Distributed Memory Model of Paragraph Vectors (PV-DM), which is analogous to the CBOW model, a paragraph (or other chosen word group) vector is added as a predictor to the context words’ vectors. The paragraph vector ”remembers” information about the paragraph beyond the selected context words and thus helps to predict the target word. The Distributed Bag of Words version of Paragraph Vectors (PV-DBOW) is analogous to the skip-gram model and uses only the paragraph vector to predict context words from the same paragraph. PV-DM provides the advantage of accounting for the sequence of the words in the paragraph, while PV-DBOW is less computationally intensive. Both, which are collectively known as doc2vec, allow efficient learning of paragraph vectors, even though different paragraphs may vary in length.
Given the success in modeling language, embeddings have been adapted to other domains, including healthcare. Researchers have used embeddings created from electronic health records (EHRs), claims data, and published biomedical literature to represent biomedical concepts, billing codes, patient visits to providers, and patients’ medical histories (Beam et al., 2018; Rajkomar et al., 2018; Miotto et al., 2016; Choi et al., 2016a; Bajor et al., 2018; Pham et al., 2017; Zhang et al., 2018; De Vine et al., 2014; Choi et al., 2016b). The last one, medical histories, are especially complex to model without embeddings because they subsume the other three; a complete and accurate representation would require modeling the similarities and differences among biomedical concepts, how they map to medical codes, and the irregular time intervals between clinical visits. Traditionally, such modeling would require slow, expensive manual crafting by biomedical and statistical experts. Miotto et al. (Miotto et al., 2016) and Bajor et al. (Bajor et al., 2018)
circumvented this slow process by using autoencoders and thedoc2vec models, respectively, to encode medical histories from EHRs. Our approach is most similar to that of Bajor et al. (Bajor et al., 2018), except that we create our embeddings from claims data, instead of from EHRs.
3. Patient-level Embeddings
3.1. Source of Data
We used de-identified claims data provided by a medium-sized, regional U.S. health insurance plan over the period of 2015-2016. The data contains demographic and enrollment information, cost, diagnostic and procedure codes, and medications. We included patients who had (1) at least one month enrollment in both years; (2) at least one medical or pharmacy claims in both years; and (3) risk scores in 2016 from a commercial risk adjustment model. A total of 441,271 patients were included in this study.
To train embeddings, we extracted all diagnoses (International Classification of Diseases [ICD]-9 or ICD-10), procedure (Current Procedural Terminology [CPT]) codes, and medications (National Drug Code [NDC]), along with their associated date stamps from each patient’s record. No data preprocessing was involved: these codes were used as-is. For each patient, we ordered the codes chronologically.
3.2. Embedding Training
We treated a patient’s entire record as a ”document,” and codes in the patient’s record as ”words” in the document. For each patient’s record, we computed an embedded representation using both doc2vec models (PV-DBOW and PV-DM). For each model, we trained combinations of embedding dimensions (100, 200, 300), sliding window sizes (10, 15, 20) with negative sampling. We used 159,457,590 codes from 441,271 patients for embedding training. All models were generated using Gensim in Python.
We conducted all our experiments using the same data source as the one used for embedding training.
4.1. Prediction Task
The prediction task was to predict prospective risk scores in 2016 using information from 2015. To calculate the risk scores in 2016, we first summed the total allowed costs in 2016. We then weighted each patient’s total cost by the total length of enrollment in 2016. Lastly, we rescaled the weighted costs to have a mean of 1.0 over the population (Hileman and Steele, 2007).
4.2. Baseline Models
4.2.1. Baseline Models 1
We developed our own risk adjustment model. We designed 21 features (Table 1) from 2015 data. These included age, sex, clinical characteristics measured using diagnostic and procedure codes, medications, healthcare utilization, and total cost. We also included a community-level description of race based on each patient’ residential ZIP code.
4.2.2. Baseline Models 2
We used a commercial risk adjustment model as another baseline model. This simple linear-additive model includes age, sex, and 150 medical markers and pharmacy markers from 2015 as features to calculate risk scores for 2016 (Hileman and Steele, 2007).
4.3. Embeddings-based Model
We used the trained embedded representations as input features for the prediction task.
4.4. Experimental Design
We evaluated the embedded representations by assessing how well they perform as input features for predicting risk scores in 2016, comparing them to the performance of the baseline models.
For the embeddings-based models, we performed a grid search over parameters (model, embedding dimension, and window size) using RIDGE regression optimized by cross-validation. We chose the best embedding parameter setting (PV-DBOW with an embedding dimension of 100 and a window size of 15) and performed the subsequent experiments using this best parameter setting.
To be consistent with previous studies in risk adjustment (Hileman and Steele, 2007), we computed three measures to evaluate the predictive accuracy: (1) ; (2) mean absolute error (MAE); and (3) predictive ratios (PRs). A predictive ratio evaluates the predictive fit at a group level and is defined as the mean predicted risk score divided by the mean actual cost for a subgroup of individuals from the sample population, with both values scaled to 1.0 over the entire population. A PR closest to 1.0 indicate a very good fit for a particular group. A predictive ratio indicates that a model overestimates the risk level for that group, whereas a PR indicates that the model underestimates the risk level. We calculated predictive ratios by age and sex groups.
We implemented all the experiments with Python. We split the dataset into training and test sets.
4.5. Prediction Results
Table 1 summarizes the patients’ characteristics in the training and test sets.
|Characteristics||Training Set||Test Set|
|Age in 2015, year||41.0 [20.0, 56.0]||41.0 [20.0, 56.0]|
|Female, n (%)||167,271 (54.2)||71,886 (54.3)|
|% of population that is African American||3.4 [2.7, 7.5]||3.4 [2.7, 7.5]|
|(based on 3-digit ZIP codes)|
|Charlson Comorbidity Index||0 [0, 1]||0 [0, 1]|
|# of claims for inpatient visits in 2015||0 [0, 0]||0 [0, 0]|
|# of claims for outpatient visits in 2015||6 [3, 13]||6 [3, 13]|
|# of claims for emergency department||0 [0, 0]||0 [0, 0]|
|visits in 2015|
|# of pharmacy claims in 2015||3 [0, 19]||3 [0, 20]|
|# of claims for specialty drugs in 2015||0 [0, 0]||0 [0, 0]|
|# of distinct drug classes in 2015||1 [0, 4]||1 [0, 4]|
|Chemotherapy in 2015, n (%)||40,936 (13.3)||17,621 (13.3)|
|Psychotherapy in 2015, n (%)||24,323 (7.9)||10,396 (7.9)|
|Obesity in 2015, n (%)||59,010 (19.1)||25,395 (19.2)|
|Cardiovascular disease in 2015, n (%)||29,002 (9.5)||12,531 (9.5)|
|Hypertension in 2015, n (%)||83,836 (27.1)||36,197 (27.3)|
|Type II diabetes in 2015, n (%)||28,424 (9.2)||12,334 (9.3)|
|Mental disorders in 2015, n (%)||106,797 (34.6)||46,029 (34.8)|
|Drug/alcohol abuse in 2015, n (%)||7,916 (2.6)||3,385 (2.6)|
|Low back pain in 2015, n (%)||47,587 (15.4)||20,467 (15.5)|
|Asthma in 2015, n (%)||32,414 (10.5)||13,826 (10.4)|
|Cost in 2015||1,591.1 [547.9, 4,909.0]||1,601.5 [555.4, 4,961.1]|
|Cost in 2016||1,634.6 [556.2, 5,133.3]||1,636.8 [556.3, 5,087.3]|
Table 2 shows the experimental results for individual-level predictive fit of three different approaches. The embeddings-based model greatly outperformed Baseline Model 2 and had a comparable performance to Baseline Model 1. For both Baseline Model 1 and the embeddings-based model, non-linear algorithms (XGBoost) had better performance than simple, linear algorithms (RIDGE).
|Baseline Model 1|
|Baseline Model 2||0.04||1.07|
Table 3 demonstrates the predictive ratios for group-level fit. Due to space limitation, we only present results from XGBoost. The embeddings-based model overestimated the risk for the elderly population with more evident bias among female and underestimated the risk for children aged 0-2. Overall, the embeddings-based model showed less evident bias as compared to Baseline Model 2.
|Baseline Model 1||Baseline Model 2||Embeddings-based Model|
In this paper, we proposed to use patient-level semantic embeddings for plan payment risk adjustment. In the experiments, the embeddings-based model showed improved performance in both individual- and group-level predictive fit compared to a commercial risk adjustment tool. Our method provided a rapid and easy-to-implement approach for risk adjustment that did not heavily rely on domain expertise, nor did it require extensive data preprocessing. A key advantage of our method was that we did not need to specify which combinations of which potential features to include in the model; instead, our approach was able to learn representations of the key factors and interactions from the claims data itself. Additionally, our method may reduce opportunities for gaming the risk adjustment system (Ellis et al., 2018; Rose, 2016), such as ”upcoding” where extra diagnostic codes or codes representing the most costly condition that a member might have are recorded (Schone and Brown, 2013). We also showed that the non-linear algorithm predicted risk scores more accurately than a linear approach. Considering the large number of individuals enrolled in insurance programs using risk adjustment, the cost-saving implications of improved risk adjustment are immense (Rose, 2016). Although our approach to risk adjustment requires further empirical evaluation, our initial work with machine learning techniques is promising.
Significant prior research has focused on developing patient representations from EHRs. However, very few studies (Choi et al., 2016b) have primarily focused on claims data, which are widely available to both providers and payers. Unlike EHRs, claims data does not include clinical notes, laboratory test results, or vital signs. However, it does encompass every interaction that a patient has with any healthcare provider, as long as a billing claim is generated. We have demonstrated the effectiveness of semantic embeddings developed solely based on claims data in a common prediction task in the health insurance domain. Given the generic nature of the embedded representation, our approach can be easily applied to a wide variety of prediction problems. In addition, future work using claims data can consider using pre-trained distributed representations learned from multimodal medical data (Beam et al., 2018), such as using claims, EHRs, and medical journals, to see if such representations will further improve the performance.
- Bajor et al. (2018) Jacek M Bajor, Diego A Mesa, Travis J Osterman, and Thomas A Lasko. 2018. Embedding Complexity In the Data Representation Instead of In the Model: A Case Study Using Heterogeneous Medical Data. (Feb. 2018). arXiv:stat.AP/1802.04233
- Beam et al. (2018) Andrew L Beam, Benjamin Kompa, Inbar Fried, Nathan P Palmer, Xu Shi, Tianxi Cai, and Isaac S Kohane. 2018. Clinical Concept Embeddings Learned from Massive Sources of Multimodal Medical Data. (April 2018). arXiv:cs.CL/1804.01486
- Bengio et al. (2003) Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Jauvin. 2003. A Neural Probabilistic Language Model. J. Mach. Learn. Res. 3, Feb (2003), 1137–1155.
- Chang and Weiner (2010) Hsien-Yen Chang and Jonathan P Weiner. 2010. An in-depth assessment of a diagnosis-based risk adjustment model based on national health insurance claims: the application of the Johns Hopkins Adjusted Clinical Group case-mix system in Taiwan. BMC Med. 8 (Jan. 2010), 7.
- Chen and Guestrin (2016) Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22Nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’16). ACM, New York, NY, USA, 785–794.
- Choi et al. (2016a) Edward Choi, Mohammad Taha Bahadori, Elizabeth Searles, Catherine Coffey, and Jimeng Sun. 2016a. Multi-layer Representation Learning for Medical Concepts. (Feb. 2016). arXiv:cs.LG/1602.05568
- Choi et al. (2016b) Youngduck Choi, Chill Yi-I Chiu, and David Sontag. 2016b. Learning low-dimensional representations of medical concepts. AMIA Summits on Translational Science Proceedings 2016 (2016), 41.
- De Vine et al. (2014) Lance De Vine, Guido Zuccon, Bevan Koopman, Laurianne Sitbon, and Peter Bruza. 2014. Medical semantic similarity with a neural language model. In Proceedings of the 23rd ACM international conference on conference on information and knowledge management. ACM, 1819–1822.
- Ellis et al. (2018) Randall P Ellis, Bruno Martins, and Sherri Rose. 2018. Chapter 3 - Risk Adjustment for Health Plan Payment. In Risk Adjustment, Risk Sharing and Premium Regulation in Health Insurance Markets, Thomas G McGuire and Richard C van Kleef (Eds.). Academic Press, 55–104.
- Hileman et al. (2016) Geoffrey R Hileman, Syed Muzayan Mehmud, and Marjorie A Rosenberg. 2016. Risk Scoring in Health Insurance: A Primer. Technical Report. Society of Actuaries, Chicago, IL.
- Hileman and Steele (2007) Geoffrey R Hileman and Spenser Steele. 2007. Accuracy of Claims-Based Risk Scoring Models. Technical Report. Society of Actuaries, Chicago, IL.
- Le and Mikolov (2014) Quoc V Le and Tomas Mikolov. 2014. Distributed Representations of Sentences and Documents. (May 2014). arXiv:cs.CL/1405.4053
- Mikolov et al. (2013a) Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013a. Efficient Estimation of Word Representations in Vector Space. (Jan. 2013). arXiv:cs.CL/1301.3781
- Mikolov et al. (2013b) Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013b. Distributed Representations of Words and Phrases and their Compositionality. In Advances in Neural Information Processing Systems 26, C J C Burges, L Bottou, M Welling, Z Ghahramani, and K Q Weinberger (Eds.). Curran Associates, Inc., 3111–3119.
- Miotto et al. (2016) Riccardo Miotto, Li Li, Brian A Kidd, and Joel T Dudley. 2016. Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records. Sci. Rep. 6 (May 2016), 26094.
et al. (2017)
Trang Pham, Truyen Tran,
Dinh Phung, and Svetha Venkatesh.
Predicting healthcare trajectories from medical records: A deep learning approach.J. Biomed. Inform. 69 (May 2017), 218–229.
- Rajkomar et al. (2018) Alvin Rajkomar, Eyal Oren, Kai Chen, Andrew M Dai, Nissan Hajaj, Michaela Hardt, Peter J Liu, Xiaobing Liu, Jake Marcus, Mimi Sun, Patrik Sundberg, Hector Yee, Kun Zhang, Yi Zhang, Gerardo Flores, Gavin E Duggan, Jamie Irvine, Quoc Le, Kurt Litsch, Alexander Mossin, Justin Tansuwan, De Wang, James Wexler, Jimbo Wilson, Dana Ludwig, Samuel L Volchenboum, Katherine Chou, Michael Pearson, Srinivasan Madabushi, Nigam H Shah, Atul J Butte, Michael D Howell, Claire Cui, Greg S Corrado, and Jeffrey Dean. 2018. Scalable and accurate deep learning with electronic health records. NPJ Digit. Med. 1, 1 (May 2018), 18.
- Rose (2016) Sherri Rose. 2016. A Machine Learning Framework for Plan Payment Risk Adjustment. Health Serv. Res. 51, 6 (Dec. 2016), 2358–2374.
- Schone and Brown (2013) Eric Schone and Randall Brown. 2013. Risk Adjustment: What is the Current State of the Art and How Can it be Improved? Technical Report. Wood Johnson Foundation, Princeton, NJ.
- Zhang et al. (2018) Jinghe Zhang, Kamran Kowsari, James H Harrison, Jennifer M Lobo, and Laura E Barnes. 2018. Patient2Vec: A Personalized Interpretable Deep Representation of the Longitudinal Electronic Health Record. (Oct. 2018). arXiv:q-bio.QM/1810.04793