Prediction of therapeutic outcomes of female pattern hair loss patients based on clinical features with application of artificial intelligence

To the Editor: Female pattern hair loss (FPHL) is common and challenging to manage. To date there is no evidence-based quantitative prediction of therapeutic response in FPHL. Artificial intelligence, particularly machine learning (ML) using algorithms to learn from provided example data in the form of features, has been used in medicine for data-driven prediction and decision making.1Rajkomar A. Dean J. Kohane I. Machine learning in medicine.N Engl J Med. 2019; 380: 1347-1358Crossref PubMed Scopus (1340) Google Scholar We aimed to develop ML models, specifically XGBoost and CatBoost algorithms, to predict therapeutic outcomes in FPHL patients at various follow-up times. XGBoost and CatBoost are gradient boosting ML algorithms with proven performance and accuracy on prediction of structured data.2Chen T. Guestrin C. XGBoost. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.2016: 785-794Google Scholar,3Kong S.H. Ahn D. Kim B.R. et al.A novel fracture prediction model using machine learning in a community-based cohort.JBMR Plus. 2020; 4: e10337Crossref PubMed Scopus (35) Google Scholar We collected data from FPHL patients with trichoscopic assessments who had received common treatment regimens from January 1, 2007 to September 30, 2019. Changes in hair density (HD) and hair caliber (HC) measured using trichoscopy at 6, 12, 18, and 24 months with a leniency of 1 month were used as therapeutic outcomes. Clinical variables and treatment regimens were included as features in model development. The process was divided into 3 steps:1.Data processing and feature selection: Variables with more than 10% missing values were excluded. Missing values were replaced by the median value. Variance inflation factor test was performed to exclude variables of multicollinearity. Pearson correlation test found no significant correlation among variables.2.Model training: Hyperparameter optimization was done by stratified 10-fold cross validation. The per-patient datasets were randomly split into the training and testing sets in an 80:20 ratio. Models were first trained on the training set on select features.3.Evaluation: Performance of predictive accuracy was evaluated using mean squared error (MSE) on the testing dataset between the observed and predicted HD and HC. Shapley additive explanation values were used to rank feature importance and to interpret the models, in particular the negative or positive impact of each feature on predictions. A total of 591 female subjects with 1918 visits were included. Average age of onset was 37.7 ± 18.0 years. Average FPHL duration before the initial visit was 78.3 ± 101 months. Nine features were selected to train ML algorithms. XGBoost and CatBoost achieved MSE ranging from 0.04 to 0.08 in prediction, indicating good accuracy. The MSE and feature importance ranking in both models are represented in Table I. A real-life example of XGBoost predication of HD change after 12 months (Fig 1) shows how it makes an individual prediction.Table IThe MSE and the feature importance ranking on HD and HC at various follow-up periods (9 clinical features were used for ML development)XGBoostCatBoostMSE∗Similarly low MSE values that were nearly 0 achieved in both models indicated their high prediction accuracy. Age of onset and FPHL duration had the greatest impacts on the prediction models.Ranking in descending order of importance†Clinical variables: Age_of_onset: age of FPHL onset; Duration: FPHL duration; FH_AGA: family history of androgenetic alopecia; TE: coexisting with telogen effluvium.‡Five common treatment regimens: oral finasteride (F), 2.5 mg to 5.0 mg once daily; 5% topical minoxidil solution or foam (Tm) once to twice daily; oral spironolactone (S), starting at 50 mg twice daily up to 100 mg twice daily; oral minoxidil (Om) 1.25 mg to 2.5 daily; and platelet-rich plasma (PRP) injections.MSE∗Similarly low MSE values that were nearly 0 achieved in both models indicated their high prediction accuracy. Age of onset and FPHL duration had the greatest impacts on the prediction models.Ranking in descending order of importance†Clinical variables: Age_of_onset: age of FPHL onset; Duration: FPHL duration; FH_AGA: family history of androgenetic alopecia; TE: coexisting with telogen effluvium.‡Five common treatment regimens: oral finasteride (F), 2.5 mg to 5.0 mg once daily; 5% topical minoxidil solution or foam (Tm) once to twice daily; oral spironolactone (S), starting at 50 mg twice daily up to 100 mg twice daily; oral minoxidil (Om) 1.25 mg to 2.5 daily; and platelet-rich plasma (PRP) injections.6 mo HD0.04Age_of_onset, Duration, TE, F, PRP, S, FH_AGA, Om and Tm0.04Duration, Age_of_onset, PRP, TE, F, Om, S, FH_AGA and Tm HC0.05Age_of_onset, Duration, F, S, FH_AGA, TE, PRP, Om and Tm0.05Age_of_onset, Duration, TE, F, S, Om, FH_AGA, Tm and PRP12 mo HD0.05Age_of_onset, Duration, TE, F, PRP, S, FH_AGA, Om and Tm0.05Age_of_onset, Duration, TE, S, FH_AGA, F, Tm, PRP and Om HC0.06Duration, Age_of_onset, S, TE, F, FH_AGA, PRP, Om and Tm0.06S, Duration, Age_of_onset, TE, FH_AGA, F, PRP, Om, and Tm18 mo HD0.06Age_of_onset, Duration, S, TE, FH_AGA, F, Tm, PRP and Om0.05S, Age_of_onset, Duration, TE, FH_AGA, Tm, PRP, Om and F HC0.08Age_of_onset, Duration, S, TE, FH_AGA, F, Tm, PRP and Om0.08Age_of_onset, Duration, S, TE, FH_AGA, F, Tm, PRP and Om24 mo HD0.05Age_of_onset, Duration, S, F, TE, FH_AGA, PRP, Om and Tm0.04Age_of_onset, Duration, S, TE, F, FH_AGA, PRP, Om and Tm HC0.04Duration, Age_of_onset, S, TE, F, FH_AGA, PRP, Om and Tm0.04Age_of_onset, Duration, S, TE, F, FH_AGA, PRP, Om and TmF, Oral finasteride; Om, oral minoxidil; PRP, platelet-rich plasma; S, spironolactone; Tm, topical minoxidil solution or foam.∗ Similarly low MSE values that were nearly 0 achieved in both models indicated their high prediction accuracy. Age of onset and FPHL duration had the greatest impacts on the prediction models.† Clinical variables: Age_of_onset: age of FPHL onset; Duration: FPHL duration; FH_AGA: family history of androgenetic alopecia; TE: coexisting with telogen effluvium.‡ Five common treatment regimens: oral finasteride (F), 2.5 mg to 5.0 mg once daily; 5% topical minoxidil solution or foam (Tm) once to twice daily; oral spironolactone (S), starting at 50 mg twice daily up to 100 mg twice daily; oral minoxidil (Om) 1.25 mg to 2.5 daily; and platelet-rich plasma (PRP) injections. Open table in a new tab F, Oral finasteride; Om, oral minoxidil; PRP, platelet-rich plasma; S, spironolactone; Tm, topical minoxidil solution or foam. To the best of our knowledge, this is the first study demonstrating the potential of ML in predicting hair growth and in determining the impacts of clinical features on clinical outcomes in FPHL patients. This is a step toward optimized and personalized treatment for FPHL patients. It revealed that age of onset and FPHL duration had greater contributions to our models, consistent with their important roles in treatment response observed in previous studies.4Won Y.Y. Lew B.L. Sim W.Y. Clinical efficacy of oral administration of finasteride at a dose of 2.5 mg/day in women with female pattern hair loss.Dermatol Ther. 2018; 31: e12588Crossref PubMed Scopus (19) Google Scholar,5Tsuboi R. Niiyama S. Irisawa R. Harada K. Nakazawa Y. Kishimoto J. Autologous cell-based therapy for male and female pattern hair loss using dermal sheath cup cells: a randomized placebo-controlled double-blinded dose-finding clinical study.J Am Acad Dermatol. 2020; 83: 109-116Abstract Full Text Full Text PDF PubMed Scopus (21) Google Scholar Because this was a retrospective study, some clinical variables, such as ethnicity, could not be incorporated into our model because of lack of documentation. Further validation studies in various populations are necessary for model improvement. Drs Shapiro and Lo Sicco are investigators for RegenLab. The other authors have no conflicts of interest to disclose.